Menu

School of Electronic Engineering and Computer Science

phd menu

PhD project ideas

Research group: Antennas and Electromagnetics
PhD title Supervisor
Research group: Centre for Digital Music
PhD title Supervisor
Research group: Cognitive Science
PhD title Supervisor
Research group: Networks
PhD title Supervisor
Research group: Risk & Information Management
PhD title Supervisor
Research group: Theory
PhD title Supervisor
Research group: Vision
PhD title Supervisor
Return to top

Improving the life-cycle of Electrical Vehicle Battery Packs using wireless or antenna networks


Research group: Antennas and Electromagnetics

Predicting the life cycle of battery pack system in Plug-in Hybrid Electrical Vehicles (PHEVs) has been the subject of studies toward the large-scale use of Electric Vehicles. With the current technology, a single battery cannot generate the performance we expect for a commercial EV. A group of batteries are used in a parallel, series or matrix form famously called as battery pack to provide the required power. Life cycle prediction of a single Li-ion battery cell is challenging due to the complexity of electrochemical modelling and thermal management. The interconnection of batteries makes the prediction more challenging as the electrical dynamics and thermal characteristics of each battery cell is different from the other cells. This may introduce random variability and the fact that aging of a single cell can propagate and reduce the life of the whole battery pack. There have been numerous approaches suggested in the literature to supress such aging propagation. This project investigates the use of wireless or antenna system in sending the State-of-the-charge (SOH), Internal resistance, Voltage and Thermal status of Li-ion battery cells. Then an integrated system should be designed to pinpoint the failing battery cell and judicially reconfigure the battery system to significantly enhance its life cycle.

Return to top

Energy storage system interface for Vehicle-to-grid (V2G) application


Research group: Antennas and Electromagnetics

The use of Electrical Vehicle (EV) battery pack to supply power in the grid (V2G) increases reliability and consistency in the grid as the renewable source, e.g. wind power, solar panel, undergoes its natural fluctuations. Furthermore, power quality can be increased with having battery storage for charging and discharging electricity to the grid. V2G operation is generally using power electronic converters (dc-dc & VSC) and inverters to act as a bidirectional charger capable of charging and discharging the battery on demand, while complying with grid standards. Commercial bidirectional chargers typically use conventional 2‐level silicon-based PWM converter topologies able to switch at relatively low frequencies. As a result, compared to the size of the battery or EV, they are relatively bulky and suffering from significant power losses. This project is focused on modelling and designing more efficient power converters to reduce the size of bidirectional chargers and reduce the power losses. This is investigated via developing novel converter topologies and control strategies for the rapid response (low latency with high switching frequency) to the grid demand.

Return to top

On-board calibration of Spark-Ignition IC engines for emission reduction


Research group: Antennas and Electromagnetics

The important process of calibration of IC engines is time-consuming and very costly for automotive companies. It costs around £1M and 18 months (in a study by Jaguar Land Rover) of hundreds of engineers using thousands of maps to calibrate a new engine and make it ready to comply with emission constraints. This project is about investigation of using a novel network of wireless sensors or actuators for on-board calibration of engines. We mean by ‘on-board’ that the calibration is done in a completely automated way while engine is running. Collected data from the wireless sensors and actuators, e.g. temperature, pressure, throttle position, should be classified and processed intelligently. Wireless sensors/actuators network must be designed in a way that it operates in harsh and noisy environment of IC engines.

Return to top

Molecular Communications for Nano-Devices


Research group: Antennas and Electromagnetics

Nanotechnology enables the miniaturization and fabrication of devices in a scale ranging from 1 to 100 nano-meters, which promises novel solutions for wide applications in biomedical (drug delivery), industrial (pollution control) and military fields (covert embedded surveillance). Although the function of a single nano-device is extremely limited, nano-networks (swarms of nano-devices) will overcome the functional limitations and potentially expand the envisaged applications. However, before the promise of nanotechnology can be fully achieved, some problems such as efficient communication between nano-devices need to be solved. The communication capabilities of nano-devices are challenging due to the small dimension, low-complexity, and low-energy expenditure constraints. Therefore, molecular communications (MC) is of particular interest to researchers because of the low-power requirements and the advantages of signal energy endurance and random propagation. MC is broadly defined as embedding digital information onto the property of molecules and allowing the molecules to diffuse from a transmitter to a set of receivers. The information can be embedded on either physical or chemical attributes of molecules. This project will primarily attempt to understand the different advantages of molecular communications in different complex biological and industrial environments, and design effective communication systems for robust transmission of data.

Return to top

Extension of QMUL-pioneered integration of semi-conducting organic polymers with antenna systems for agile, optical, front-end control of beam patterns


Research group: Antennas and Electromagnetics

This thesis topic builds on recent heritage pioneered at QMUL between the School of Physics & Astronomy and EECS to integrate semi-conducting organic polymers into antenna systems to open up new capabilities for beam-pattern formation and steering. Resultant materials and methods are to be integrated into commercial terrestrial wireless networks for more robust operation in extremes of temperature.

Return to top

Application of computational electromagnetic modelling via CST to guide clinical interpretation of images acquired by Optical Coherence Tomography


Research group: Antennas and Electromagnetics

This topic concerns computer-assisted analysis of OCT images to aid dental clinicians in initial screenings for oral cancer. The aim is to provide ‘computer-vision’ that guides more certain diagnosis. Underpinning this is the need to understand mechanisms of scattering of a probing light-beam. The electromagnetic modelling package CST will be employed to do this by various canonical studies of beam-field scattering. This thesis will be undertaken in collaboration with Dr. Peter Tomlins of the Institute of Dentistry, Queen Mary University of London.

Return to top

Nano-scale EM/Hybrid Communication Networks for Healthcare Monitoring Applications


Research group: Antennas and Electromagnetics

With the development of nanotechnology, the idea of connecting nano-devices to conduct complicated tasks and communicate the information collected by these devices was a natural progression in order to complete the overall picture of a new generation of connected devices. As a consequence, nano-networks were proposed by the IEEE standardization group (P1906.1 - Recommended Practice for Nano-scale and Molecular Communication Framework, which the principle supervisor is a member of) and therefore the need for nano-communication was a necessity. Nano-networking is the study of communication among devices and/or entities – manmade, biological, and hybrid – with very small dimensions; challenging physical features of this communication environment make analysis and system design very different from conventional communication systems. Nano-networking is a rapidly emerging discipline, but as yet (1) emerging technology trends and important open problems in this area are unclear; and (2) new industrial applications need to be identified.

Return to top

Reconfigurable Radio Front-ends from UHF to Mm-wave for Future Wireless Communications


Research group: Antennas and Electromagnetics

Cognitive Radio is one of the potential wireless applications that may place severe demands on RF system designers and particularly antenna designers, when it comes to providing flexible radio front-ends capable of achieving the set objectives of the technology. The aim of this project is to investigate possible roles that different categories of reconfigurable antenna can play in Cognitive and smart Radio. Hence, our research focuses on investigating some novel methods to frequency-reconfigure compact ultra-wideband antennas to work in different bands; this will offer additional filtering to the radio front-end. In the ultra-wideband mode, the antenna senses the spectrum for available bands with less congestion and interference and, hence, decides on the most suitable part to use at that point, initiating the necessary reconfiguration, allowing reliable and efficient communication links between the radio devices. Furthermore, the design of novel pattern and polarisation reconfigurable antennas is to be investigated to assist Cognitive Radio through spatial, rather than frequency, means.​​

Return to top

Exploration into the Development of Miniaturised Electromagnetic Wave Detection Methods for Medical Diagnostics


Research group: Antennas and Electromagnetics

The interaction of Electromagnetic (EM) waves with the human body has been the subject of numerous studies, and has resulted in the development of various diagnostic and treatment applications. This project aims to explore into the development of advanced miniaturised and wearable diagnostic EM systems for monitoring the dynamic impedance and other parameters within the human body. The developed technology could provide information on a number of parameters which could be used as diagnostic tools for monitoring the human body condition.

Return to top

Euler-Lagrangian modelling of THz energy modes of ensemble oscillation in protein systems


Research group: Antennas and Electromagnetics

The challenge of this topic is to assign labels of mechanical action to measured THz transmission spectra. A corollary of this topic is to up-grade the experimental capability of the THz laboratory to undertake high-quality 2D time-domain spectroscopy in order to more finely resolve spectral information. The capacity to cleanly measure THz energy action in bio-molecular systems, and understand the origins of the spectra, is fundamental to elucidating mechanisms of disease and potential drug development.

Return to top

Data collection and analysis on UHF white space spectrum and beyond


Research group: Antennas and Electromagnetics

Spectrum measurement has become increasingly important for radio environment map which enables the cognitive radio into practice. Two schemes have been proposed to determine the spectrum occupancy. One is geo-location which is based on a central database to indicate the maximum permitted power for transmission in each vacant channel at a specific location and certain period. Another scheme is spectrum sensing, which detects spectrum holes that are not occupied by primary users. In this project, we would like to have a candidate who will be able to use our exiting spectrum monitoring device and the geo-location databased qualified by the UK Ofcom to collect the spectrum data from UHF whites space to 6GHz, then use data analytics and machine learning to derive the insights for future wireless communications.

Return to top

Novel Control methodologies for emission reduction in internal-combustion SI petrol or diesel engines


Research group: Antennas and Electromagnetics

Reduction of CO2 emission and other particulates produced from internal-combustion engine is still a challenge. The recent fiasco of Volkswagen in cheating the emission production data of their diesel engines (http://www.theguardian.com/business/2015/sep/22/vw-scandal-caused-nearly-1m-tonnes-of-extra-pollution-analysis-shows) proves that we still need smarter calibration/control techniques to restrict the emission within the EU standards. To achieve this, we need more accurate real-time 1D or 0D models of engine to predict emission in realistic scenarios. Then these models can be used for the investigation of novel model-based control strategies to reduce the emission. 2 or 3 projects can be defined to focus on modelling, control system development and downsizing Spark-ignition (SI) petrol engine or diesel engine. I have a close collaboration with the Control Systems research group of University of Manchester on engine control development. I am also open to any new ideas on the emission reduction in IC engines if PhD candidates would like to propose.

Return to top

Development of Novel RF Fingerprinting Location-Detection System for indoor applications


Research group: Antennas and Electromagnetics

By considering the inverse of the traditional triangulation systems, a single RFID reader (or more) could be used in conjunction with location-referenced RFID tags to help locate an RFID tagged target based on comparing the different responses. This approach is also known as RF fingerprinting since it estimates the target’s location by matching its measured RF responses with the previously collected RF fingerprints of the reference tags (Typically RSS values). This project aims at applying improved LANDMARC and VIRE algorithms to resolve the RF fingerprint for locating targets within indoor environments. The project also focuses on a number of other aspects such as the development of suitable directional antennas to complement the optimised fingerprinting algorithms. These improvements aim to limit the searching space, attain higher gain per element, and minimise the multipath effects in such complex propagation environments.

Return to top

Exploration into the Development of Miniaturised Electromagnetic Wave Detection Methods for Medical Diagnostics


Research group: Antennas and Electromagnetics

The interaction of Electromagnetic (EM) waves with the human body has been the subject of numerous studies, and has resulted in the development of various diagnostic and treatment applications. This project aims to explore into the development of advanced miniaturised and wearable diagnostic EM systems for monitoring the dynamic impedance and other parameters within the human body. The developed technology could provide information on a number of parameters which could be used as diagnostic tools for monitoring the human body condition.

Return to top

Wireless energy measurement in home environment


Research group: Antennas and Electromagnetics

Low cost monitoring of energy usage in home is still a challenging problem and considering the rise of energy cost is very important for any household. This project is focused on designing and implementing a wireless network so that the user can monitor the usage of any home appliances in the home environment. Classifying the pattern of energy consumption of a device in real-time in a wireless network of home appliances is by far difficult and can be considered as a machine-learning problem. As a further step, using the energy usage pattern data, an intelligent system can be implemented to learn the behaviour of home users to reduce the cost of energy. This system would remotely control the power consumption of each appliance instead of human in the network and judicially switch on/off devices. There are many challenging and interesting problems in this project including calculating the real power usage (RMS) of each home appliance, recognising the power usage pattern of individual home appliances in the aggregated received patterns from wireless devices, designing the suitable wireless device (RFID or wireless sensors/actuators), an intelligent system to make decisions in real-time to reduce the cost of household energy usage.

Return to top

Development of Novel RF Fingerprinting Location-Detection System for indoor applications


Research group: Antennas and Electromagnetics

By considering the inverse of the traditional triangulation systems, a single RFID reader (or more) could be used in conjunction with location-referenced RFID tags to help locate an RFID tagged target based on comparing the different responses. This approach is also known as RF fingerprinting since it estimates the target’s location by matching its measured RF responses with the previously collected RF fingerprints of the reference tags (Typically RSS values). This project aims at applying improved LANDMARC and VIRE algorithms to resolve the RF fingerprint for locating targets within indoor environments. The project also focuses on a number of other aspects such as the development of suitable directional antennas to complement the optimised fingerprinting algorithms. These improvements aim to limit the searching space, attain higher gain per element, and minimise the multipath effects in such complex propagation environments.

Return to top

Software-defined radio for machine-to-machine spectrum challenge


Research group: Antennas and Electromagnetics

Spectrum for wireless communicants has become exponentially scarce, especially for the next generation system such as machine-to-machine and 5G. Dynamic spectrum access has been well studied theoretically, but very limited real world implementations. Both the UK regulator, Ofcom, and the US regulator, FCC, have initialised signifiant amount of efforts to enable the UHF TV bands as a starting point for the dynamic spectrum access. In this project, we would like to have a candidate who are interested in soft-defined radio platform such as NI USRP, and take our exiting compressive spectrum sensing algorithms and Geo-Location database model into the software defined radio platform, and participant in part of the Ofcom white space device study in the UK.

Return to top

Wearable computing systems for auditory display


Research group: Centre for Digital Music

Auditory display is the audio analogue of visual display. It uses sound to communicate information and may be used in a multimodal human-computer interface. This project will look at how auditory display can be implemented for wearable computing. It will combine audio signal processing, possibly incorporating spatial audio, and may look at other modes of interaction such as haptic feedback. It can also extend to exploring new designs of worn actuators (speakers) beyond current in-ear solutions and new ways of conveying complex haptic information in a body-centric system.

Return to top

Textiles as flexible substrates for circuitry


Research group: Centre for Digital Music

Electronic textiles (e-textiles) incorporate conductive fibres into a textile that acts as an electrical circuit. It is a growing field, particularly popular with the current interest in wearable tech. Current commercial products use e-textiles in sensors to measure signals such as heartbeat and breath, but the electronics used to capture these signals and send them to a gateway device such as a phone remain on hard PCBs. These PCBS are usually snapped onto the garment and have none of the material benefits of the e-textiles. This project examines the hard/soft connections between textile and electrical components. It can include, but is not limited to, techniques that combine weaving, knitting, embroidery, and print. The aim will be to reduce the fragility of the hard/soft interfaces and to move towards an electrical system fully integrated into a textile.

Return to top

Sound Event Detection in Everyday Audio


Research group: Centre for Digital Music

The emerging field of sound scene analysis refers to the development of software systems for automatically recognising everyday sounds and the environment/context of a recording. Applications of sound scene analysis include smart homes, urban planning, audio-based security/surveillance, indexing of sound archives, and acoustic ecology. This project will focus on detecting sound events from everyday sound scenes. You will carry out research and develop computational methods suitable for detecting overlapping sound events from noisy and complex audio, recorded in urban environments. In this project you will be based in the Machine Listening Lab in the Centre for Digital Music, developing new methods and software tools based on signal processing and machine learning theory.

Return to top

Recognition and Separation of Musical Instruments in Polyphonic Audio


Research group: Centre for Digital Music

One of the key problems in the field of music informatics is to develop systems that can automatically recognise musical instruments, however there is relatively little research in identifying instruments in polyphonic mixtures. The goal of this project is to develop computational techniques for automatic identification/separation of multiple instruments in complex music signals, as well as to develop methods for instrument assignment - assigning detected notes to a specific instrument. In this project you will be based in the Machine Listening Lab in the Centre for Digital Music, developing new methods and software tools based on signal processing and machine learning theory.

Return to top

Novel Virtual Reality or Augmented Reality experience for creativity


Research group: Centre for Digital Music

This project would develop novel Virtual Reality or Augmented Reality experience to support people being creative together. Multiple people would create together either co-located, or remotely. The Interaction Design would involve systematic evaluation of the engagement between people in these creative virtual or augmented experiences.

Return to top

Visualisations of Creativity and Mutual Engagement


Research group: Centre for Digital Music

Using interactive technologies to create new kinds of creativity support environments where multiple people create together at once either co-located, or remotely. The project would involve developing visualisations of the creative process and using these to systematically evaluate the engagement between people in these creative experiences.

Return to top

Experimental design for machine listening and other applications of machine learning


Research group: Centre for Digital Music

Funny this: the design of experiments -- so essential to science and industry because it helps one obtain the most reliable measurements at the least cost -- remains by and large absent from evaluation methodology in machine listening in particular, and machine learning in general. There are several reasons for this, but a major one is that we can run thousands of classification experiments to obtain thousands of numbers in little time, and then apply standard statistical tools to formally test hypotheses. Hence, it seems that experiments in machine learning are cheap enough that _design_ does not really matter. However, when we scratch below the surface of these experiments some big problems arise that cannot be ignored. Some recent work attempts to correct this, but so much more needs to be done. This research aims to contribute to building a foundation of experimental design that is compatible with the unique nature of machine learning experiments, specifically when applied to machine listening.

References

C. Drummond and N. Japkowicz, “Warning: Statistical benchmarking is addictive. kicking the habit in machine learning,” J. Experimental Theoretical Artificial Intell., vol. 22, pp. 67–80, 2010.
N. Japkowicz and M. Shah, Evaluating Learning Algorithms: A Classification Perspective. New York, NY, USA: Cambridge University Press, 2011.
E. R. Dougherty and L. A. Dalton, “Scientific knowledge is possible with small-sample classification,” EURASIP J. Bioinformatics and Systems Biology, vol. 2013:10, 2013.
K. L. Wagstaff, “Machine learning that matters,” in Proc. Int. Conf. Machine Learning, 2012.

Note: I am interested in supervising other projects related to signal processing and music informatics; please contact me to discuss.

Return to top

Intelligent Machine Listening


Research group: Centre for Digital Music

In this research, we seek ways to exploit novel and holistic approaches to evaluation for building machine listening systems (and constituent parts). A major emphasis will be on answering “how” systems work and “what” they have learned to do, in relation to the success criteria of real-world use cases. The research will involve working at the intersection of digital signal processing, machine learning, and the design and analysis of experiments.

References

K. L. Wagstaff, “Machine learning that matters,” in Proc. Int. Conf. Machine Learning, 2012.
O. Pfungst, Clever Hans (The horse of Mr. Von Osten): A contribution to experimental animal and human psychology. New York: Henry Holt, 1911.
B. L. Sturm, “A simple method to determine if a music information retrieval system is a “horse”,” IEEE Trans. Multimedia, vol. 16, no. 6, pp. 1636–1644, 2014.
B. L. Sturm, “The state of the art ten years after a state of the art: Future research in music information retrieval,” J. New Music Research, vol. 43, no. 2, pp. 147–172, 2014.

Note: I am interested in supervising other projects related to signal processing and music informatics; please contact me to discuss.

Return to top

Automatic music genre recognition


Research group: Centre for Digital Music

As a composer, I am confounded by research that aims to build a computer system that labels a “piece of music” as belonging to one of a handful of speciously defined, culturally negotiated, dynamic, and dare I say, “fetishistic,” “genres" that are supposedly split apart and redefined to suit whatever biases one may have to the tune of, e.g., “I listened to math rock before it was cool.” But enough about me and my “hangups.” This is about you. No – this is about us, versus a world where we can’t deny that music “genres" exist, and people use the vocabulary of genre to discuss, perform and make music, to organise their music collections, to make recommendations, and to discover new music (but apparently not my music). This topic has been researched a considerable amount in music informatics, but almost all of the time “genre” is defined implicitly via some “cleanly labeled” dataset. It is time to take a dramatically different approach, unifying signal processing and machine learning with use case definitions and expert domain knowledge.

References:

O. Pfungst, Clever Hans (The horse of Mr. Von Osten): A contribution to experimental animal and human psychology. New York: Henry Holt, 1911.
J. Frow, Genre. New York, NY, USA: Routledge, 2005.
F. Holt, Genre in popular music. The University of Chicago Press, 2007.
R. O. Gjerdingen and D. Perrott, “Scanning the dial: The rapid recognition of music genres,” J. New Music Research, vol. 37, pp. 93–100, Spring 2008.
B. L. Sturm, “A simple method to determine if a music information retrieval system is a “horse”,” IEEE Trans. Multimedia, vol. 16, no. 6, pp. 1636–1644, 2014.
B. L. Sturm, “The state of the art ten years after a state of the art: Future research in music information retrieval,” J. New Music Research, vol. 43, no. 2, pp. 147–172, 2014.
B. L. Sturm, “A survey of evaluation in music genre recognition,” in Adaptive Multimedia Retrieval: Semantics, Context, and Adaptation (A. Nu ̈rnberger, S. Stober, B. Larsen, and M. Detyniecki, eds.), vol. LNCS 8382, pp. 29–66, Oct. 2014.

Note: I am interested in supervising other projects related to signal processing and music informatics; please contact me to discuss.

Return to top

Sound effect synthesis


Research group: Centre for Digital Music

Sound synthesis is the generation of sounds using algorithms. It is an important application for cinema, multimedia, games and sound installations. This research topic is focused on a ‘big data’ approach to sound synthesis. It addresses the as yet unsolved problem of how to synthesize an entire sound effect library, without creating many different unique models. Learning systems will be fed entire sound effects libraries, capturing the statistics and audio features most associated with classes of sounds. This will be used to adapt the parameter settings of general purpose synthesis models such that they can be tailored to generate sounds that capture the most relevant features of samples or classes of samples in a sound effects library. There are three main challenges in this direction of research; the choice of synthesis model(s) to use, the features used to establish sound similarity, and the learning or optimization technique employed to find the best setting of parameters to ensure that the model generates a close perceptual match to a sound sample. This topic represents a novel research direction with potentially high impact. The researcher will be encouraged to present the research at high impact conferences and publish the results in premier, peer reviewed journals.

Return to top

Acoustic autofocus


Research group: Centre for Digital Music

This project will pioneer research and development of an auto-focus for audio. It builds on the proposal of an acoustical zoom using Directional Audio Coding (DirAC). However, DirAC requires B Format microphones, which are large and expensive. The researcher will investigate methods of real-time stereo field decomposition to determine the direction of arrival of sound sources. Focus may be achieved by adjusting gain, delay, reverb and equalisation of the two microphone signals in order to emphasise those sources whose direction of arrival is nearer to the centre of the stereo soundfield. Other solutions, such as microphones with variable directivity, will also be investigated. Outcomes of the research could include improved methods for real-time direction of arrival estimation, psychoacoustic studies of preference for sound scene editing and a demonstrator that includes a stereo microphone synchronized to the zoom lens of a video camera.

Return to top

Intelligent systems for audio production


Research group: Centre for Digital Music

This PhD topic aims to intelligently generate an automatic sound mix out of an unknown set of multi-channel inputs. The research also explores the possibility of creating an intelligent system that ‘learns’ the mixing decisions of a skilled audio engineer with minimal or no human interaction. The input channels can be analysed to determine preferred settings for gain, equalization, compression, reverb, and so on. Alternatively, entirely new approaches to audio production can be taken, since an intelligent system is capable of far more analysis and manipulation than can be performed manually. This research has application to live music concerts, remote mixing, recording and postproduction as well as live mixing for interactive scenes.
The justification of this research is the need of non-expert audio operators and musicians to be able to achieve a quality mix with minimal effort. Currently, mixing is a task which requires great skills, practice and can be sometime tedious. For the professional mixing engineer this kind of tool will reduce sound check time and allow the engineer to focus on the creative aspects of production.
This research topic builds on previous, successful work by researchers within the Centre for Digital Music, but is broad enough in scope that it could be taken in new and exciting directions.

Return to top

Machine-learning architectures for making sense of bird sounds


Research group: Centre for Digital Music

Natural sounds such as birdsong have rich time structure. The "information" in birdsong is not just in the "notes" that are sung, but also in the sequence, the durations, and the size of the pauses between. Yet machine learning systems often use very simple assumptions about temporal structure. How can we build machine learning systems that reflect and represent the rich time structure of birdsong? Possibilities include neural network architectures such as Long Short-Term Memory (LSTM) or Denoising Auto-Encoders, unsupervised feature learning, or hierarchical probabilistic models. How will we know success? Well, possibly by building a better bird recognition system. But further! Maybe the system can explain to us, in words or numbers, what makes a robin sound different from a blackbird. Maybe the system can learn the structure of one species' song, and then go on to sing new melodies it has composed for itself. This topic is suitable for someone with a background in a mathematical topic such as machine learning or statistical optimisation, ideally with a passion for sound.

Return to top

Birdsong similarity measured through structured temporal models


Research group: Centre for Digital Music

Birdsong is highly complex and structured. Yet, even human listeners can often make robust judgments about the relative similarity/difference of two bird song recordings, for example when distinguishing between two different bird species, or two individual birds, or two song types produced by the same individual. Animal communication researchers now have large amounts of data and so they need automatic similarity judgments to help them uncover patterns of cultural and evolutionary transmission. Automatic techniques are commonly used already (cross-correlation, dynamic time warping, Markov models) but these methods are known to ignore complexities such as hierarchical structure, and so their judgment is not rich enough to reproduce perceptual similarity as well as is needed. In this project you will be based in the Machine Listening Lab (in the Centre for Digital Music), using statistical models, machine learning and signal processing to develop a method for similarity measurement that can be applied to a wide variety of bird sounds. Methods might include hierarchical Markov models, state-space models, probabilistic grammars, semi-Markov models, recurrent neural networks (RNNs), information dynamics. You will be collaborating with Dr Rob Lachlan (School of Biological and Chemical Sciences) to work with rich scientific datasets of recorded bird sounds.

Return to top

Textile sensors for worn interfaces


Research group: Centre for Digital Music

Electronic textiles (e-textiles) incorporate conductive fibres into a textile that acts as an electrical circuit. It is a growing field, particularly popular with the current interest in wearable tech. Some work has been done around creating fabric sensors that respond to touch and even some initial work on capturing more complex gestures such as seen in Google’s Project Jacquard. This project would push forward the technology that could support more nuanced interaction design. It would ask how could the textile form inform the interaction. In other words, can the textile design and construction of a garment aid in the user interaction of a worn computing system? This project would combine sensor design and the signal processing of sensor data with craft techniques in order to build prototypes and test new sensor systems.

Return to top

Wearable computing systems for auditory display


Research group: Centre for Digital Music

Auditory display is the audio analogue of visual display. It uses sound to communicate information and may be used in a multimodal human-computer interface. This project will look at how auditory display can be implemented for wearable computing. It will combine audio signal processing, possibly incorporating spatial audio, and may look at other modes of interaction such as haptic feedback. It can also extend to exploring new designs of worn actuators (speakers) beyond current in-ear solutions and new ways of conveying complex haptic information in a body-centric system.

Return to top

Textiles as flexible substrates for circuitry


Research group: Centre for Digital Music

Electronic textiles (e-textiles) incorporate conductive fibres into a textile that acts as an electrical circuit. It is a growing field, particularly popular with the current interest in wearable tech. Current commercial products use e-textiles in sensors to measure signals such as heartbeat and breath, but the electronics used to capture these signals and send them to a gateway device such as a phone remain on hard PCBs. These PCBS are usually snapped onto the garment and have none of the material benefits of the e-textiles. This project examines the hard/soft connections between textile and electrical components. It can include, but is not limited to, techniques that combine weaving, knitting, embroidery, and print. The aim will be to reduce the fragility of the hard/soft interfaces and to move towards an electrical system fully integrated into a textile.

Return to top

Recognition and Separation of Musical Instruments in Polyphonic Audio


Research group: Centre for Digital Music

One of the key problems in the field of music informatics is to develop systems that can automatically recognise musical instruments, however there is relatively little research in identifying instruments in polyphonic mixtures. The goal of this project is to develop computational techniques for automatic identification/separation of multiple instruments in complex music signals, as well as to develop methods for instrument assignment - assigning detected notes to a specific instrument. In this project you will be based in the Machine Listening Lab in the Centre for Digital Music, developing new methods and software tools based on signal processing and machine learning theory.

Return to top

Sound Event Detection in Everyday Audio


Research group: Centre for Digital Music

The emerging field of sound scene analysis refers to the development of software systems for automatically recognising everyday sounds and the environment/context of a recording. Applications of sound scene analysis include smart homes, urban planning, audio-based security/surveillance, indexing of sound archives, and acoustic ecology. This project will focus on detecting sound events from everyday sound scenes. You will carry out research and develop computational methods suitable for detecting overlapping sound events from noisy and complex audio, recorded in urban environments. In this project you will be based in the Machine Listening Lab in the Centre for Digital Music, developing new methods and software tools based on signal processing and machine learning theory.

Return to top

Textile sensors for worn interfaces


Research group: Centre for Digital Music

Electronic textiles (e-textiles) incorporate conductive fibres into a textile that acts as an electrical circuit. It is a growing field, particularly popular with the current interest in wearable tech. Some work has been done around creating fabric sensors that respond to touch and even some initial work on capturing more complex gestures such as seen in Google’s Project Jacquard. This project would push forward the technology that could support more nuanced interaction design. It would ask how could the textile form inform the interaction. In other words, can the textile design and construction of a garment aid in the user interaction of a worn computing system? This project would combine sensor design and the signal processing of sensor data with craft techniques in order to build prototypes and test new sensor systems.

Return to top

Musical Models for Transcription of the Singing Voice


Research group: Centre for Digital Music

Research on automatic transcription has focussed mainly on the extraction of pitch and timing from audio recordings. A transcription which is useful to an end-user typically needs to incorporate a range of further information: interpreting and expressing the trajectories in the time-frequency plane as discrete events, the timing in terms of the implied metrical structure, and the pitches in terms of the implied scale. The first goal of this project is to model musical knowledge and stylistic conventions in order to produce a transcription of solo singing to Western notation. Western notation, however, is an incomplete representation of a performance, and although there are means for describing expression qualitatively, there is no accepted representation for continuous pitch trajectories, nor for timbre (e.g. phonation mode). The second goal of the project is thus to develop a representation for singing transcription that facilitates the development of algorithms for assessment of various characteristics such as similarity to a reference recording, accuracy of pitch and timing, quality of vocal sound, and the scale and intonation which are used in the singing. In the final stage of the project, these algorithms will be applied in a cross-cultural study of singing, to test the robustness of the representations and correctness of the assumptions.
Note: I am interested in supervising other projects related to music informatics; please contact me to discuss.

Return to top

Instrument modelling to aid polyphonic transcription


Research group: Centre for Digital Music

A leading current approach to transcription is based on the factorisation of a time-frequency representation into a dictionary of instrument sounds and a matrix of instrument activities over time, where the dictionary typically contains one or very few templates per pitch and instrument. One of the main weaknesses of such an approach is its failure to model the range of sounds that can be produced by an instrument, including the spectral variation which occurs even during a single instance of a note. Research in music acoustics provides detailed models of the mechanics of instruments and the resulting range of sounds that they produce, and the goal of this project is to apply this knowledge to the analysis of polyphonic mixtures of instruments, in order to achieve more accurate decompositions. The final objective is to develop a fully automatic system, which performs instrument recognition in an initial analysis stage, and then adapts the transcription algorithm on the basis of the instrument models which are relevant to the recording.
Note: I am interested in supervising other projects related to music informatics; please contact me to discuss.

Return to top

Transcribing the Ineffable in Music


Research group: Centre for Digital Music

In a collaborative art piece created by Lina Viste Grønli, Peter Child, and Elaine Chew, Chew’s first encounter with a Haydn sonata movement is meticulously transcribed by Child to create a performable score. The score, a refraction of Haydn’s original piece, captures not only the surface repetitions, errors, halts, and interruptions of the practice, but also encodes in notation the deeper expressive cues of Chew's final performance, and Child’s cognition of the performer’s practice. The goal of this project is to replicate Child’s actions using computer models so as to automatically create notations of the performer’s and the composer’s musical interpretations for this and other performances of Haydn and beyond. Backgrounds in algorithm design, software implementation, and rhythm transcription strongly recommended.

Reference:
Chew, E., Child, P., Grønli, L. V. (2013). Practicing Haydn (Piano Sonata in Eb, Hob XVI:45 finale). https://d3mdhum531ctfj.cloudfront.net/exhibitions/Lina_V_Gronli/Practicing_Haydn_Score-1.pdf
Chew, E. (2013). About Practicing Haydn. http://elainechew-piano.blogspot.co.uk/2013/09/about-practicing-haydn.html

Return to top

What Makes Music Beautiful?


Research group: Centre for Digital Music

At the 2016 Dagstuhl Workshop on Computational Music Structure Analysis, Masataka Goto threw out the grand challenge of deciphering beauty in music. This project aims to take up this challenge to formalise and understand the nature of beauty in music. Geometer H. S. M. Coxeter, who was a composer in his youth and an important influence on the art of M. C. Escher, writes on the aesthetic analogy between mathematics and music. He quotes G. H. Hardy (1940), “There is a very high degree of *unexpectedness* combined with *inevitability* and *economy* … A *mathematical proof* should resemble a simple and clear-cut constellation, not a scattered cluster in the Milky Way” and goes on to say, “Similar words might well be used as advice to composers, with ‘mathematical proof’ replaced by ‘piece of music.’” A goal of this project will be to create mathemusical expressions of Hardy’s aesthetic rules of unexpectedness, inevitability, and economy. Backgrounds in music and mathematics (or a closely related field) are strongly recommended.

References:
Coxeter, H. S. M. (1968). Music and Mathematics. The Canadian Music Journal, 61(3):312-320. http://www.jstor.org/stable/27957839
Hardy, G. H. (1940). A Mathematician’s Apology. Cambridge: Cambridge University Press.

Return to top

Recognition and Separation of Musical Instruments in Polyphonic Audio


Research group: Centre for Digital Music

One of the key problems in the field of music informatics is to develop systems that can automatically recognise musical instruments, however there is relatively little research in identifying instruments in polyphonic mixtures. The goal of this project is to develop computational techniques for automatic identification/separation of multiple instruments in complex music signals, as well as to develop methods for instrument assignment - assigning detected notes to a specific instrument. In this project you will be based in the Machine Listening Lab in the Centre for Digital Music, developing new methods and software tools based on signal processing and machine learning theory.

Return to top

Sound Event Detection in Everyday Audio


Research group: Centre for Digital Music

The emerging field of sound scene analysis refers to the development of software systems for automatically recognising everyday sounds and the environment/context of a recording. Applications of sound scene analysis include smart homes, urban planning, audio-based security/surveillance, indexing of sound archives, and acoustic ecology. This project will focus on detecting sound events from everyday sound scenes. You will carry out research and develop computational methods suitable for detecting overlapping sound events from noisy and complex audio, recorded in urban environments. In this project you will be based in the Machine Listening Lab in the Centre for Digital Music, developing new methods and software tools based on signal processing and machine learning theory.

Return to top

Deep Learning for Music Compression: Adversarial Training with Attention Modelling


Research group: Centre for Digital Music

The development of music compression was at its peak between 1990-2000. At the time, the automatic identification of perceptually irrelevant information in audio streams enabled a drastic decrease in bitrate and had a massive impact on how we listen to, buy and produce music today. Since then, however, the underlying concepts have hardly changed and can be found in essentially identical form on all subsequent audio coding standards (MPEG surround, SAOC, MPEG-H,...). In lossless audio compression, linear prediction remains a central component that has not been replaced since the 1970s. The goal of this PhD is to design deep neural networks that can learn to anticipate music signals and thus to create a conceptually novel approach for further reducing the bitrate. Such a method can (but not necessarily must) be combined with existing perceptual coding techniques. This problem scenario could involve various types of networks for temporal modelling (possibly with a specific focus on attention-based models to incorporate offline knowledge) and adversarial training (as music streams are compared to the frame-rate of the spectral front-end rather stationary processes and adversarial training helps learning the behaviour at boundaries between such processes).

Return to top

Context-Adaptive Sound Separation with Deep Networks


Research group: Centre for Digital Music

Given an audio recording of a piece of music, the goal of sound separation is to isolate and un-mix individual sounds (related to instruments, notes, melody, background) from the recording. Sound separation is often referred to as a key technology in music processing as it enables direct access to individual sound events and thus increases ways to analyse, process and re-use a recording (e.g. for re-mixing and up-mixing sounds, or for musical analysis). Unfortunately, without making prior assumptions about the recording, sound separation is mathematically ill-posed and unsolvable. However, if rich prior knowledge can be provided, the results are useful for a variety of applications. The goal of this PhD is to investigate in the context of audio sound separation, different approaches to incorporating prior knowledge into deep neural networks or to imposing informed constraints on the learning process. This way, building on recent success in deep learning, the prior knowledge can be used to constrain the expressive and analytical power provided by neural networks, resulting in potentially higher separation quality. In this context, the use and development of (Bayesian) graph networks, variational autoencoders and other methods combining probabilistical modelling with neural networks is of central interest.

Return to top

Deep Learning for Evaluation: Differentiable Metrics from the Crowd


Research group: Centre for Digital Music

Evaluation is of central importance not only in music informatics but in signal processing and machine learning in general - only with proper evaluation it is possible to state whether a proposed system is likely to do what we want it to do. Therefore, the more accurate and representative the evaluation, the more conclusive the results. For example, in music processing, listening tests are often the only reliable way to obtain perceptually relevant evaluation metrics. In machine learning, however, such evaluations have to be conducted thousands or millions of times, as the parameters of a system are adjusted to maximize an evaluation measure - here, listening tests are too time-consuming and expensive. As a makeshift, most machine learning approaches employ simplistic evaluation measures such as a Euclidean distance or a cross-entropy. The goal of this PhD is to develop neural networks that can learn to predict the results of listening tests, with training data obtained using crowd-sourcing services such as Amazon Turk or similar platforms. Such a network is intrinsically differentiable and thus is immediately useful as an objective function in machine learning and thus can directly be used to improve upon the currently used simplistic measures. In this sense, the goal of this PhD is not to replace listening tests but to improve upon how we train our models in general.

Return to top

Research group: Centre for Digital Music

While the simple, single search or "find" mechanism routinely provided in standard interfaces to wordprocessors, spreadsheets etc. is often adequate, there are times when it is desirable to perform multiple, simultaneous searches or searches for multiple terms occurring together. This project will investigate ways of implementing more sophisticated search mechanisms in an intuitive way, so that they form a natural and easily understood part of routine interaction. This is different from the development of systems specifically to support complex information seeking, as the focus is the design and integration of complex searches within every day systems and tasks. Issues to be examined include the breaking down of complex search expressions and the implementation of these in a way that is immediately apparent to users, and the processing and presentation of complex multi-criteria search results that provides a well thought through match to the users task and context. The project will involve the iterative development of multi-search interface prototypes, and result in detailed recommendations of evaluated means of implementing multi-search functions. At least two different PhD projects are possible, focusing respectively on mobile and desktop platforms.

Return to top

Sonic Exploration of Physiological Control


Research group: Centre for Digital Music

Physiological signals such as Heart Rate Variability (HRV), Finger Capillary Blood Flow (FCBF) and Respiration (r) provide a means of examining the health and reactivity of our bodily control mechanisms, such as the influence of R on HRV. Traditionally, analysis of these systems has used visualisations of individual signals or signals in combination. This project will explore what data sonification methods can bring to the analysis of physiological control systems, either in compliment to visual analysis or on its own. The specific focus of any one PhD project may be as narrow as the detailed investigation of variations in one signal, such as HRV, in response to controlled forcing frequency inputs, such as controlled breathing rates, or may focus on a specific control mechanism involving multiple signals, such as the baroreflex control of the cardiovascular system, that is relevant to the analysis and modelling of cardiovascular oscillations and regulation.

Return to top

Experimental design for machine listening and other applications of machine learning


Research group: Centre for Digital Music

Funny this: the design of experiments -- so essential to science and industry because it helps one obtain the most reliable measurements at the least cost -- remains by and large absent from evaluation methodology in machine listening in particular, and machine learning in general. There are several reasons for this, but a major one is that we can run thousands of classification experiments to obtain thousands of numbers in little time, and then apply standard statistical tools to formally test hypotheses. Hence, it seems that experiments in machine learning are cheap enough that _design_ does not really matter. However, when we scratch below the surface of these experiments some big problems arise that cannot be ignored. Some recent work attempts to correct this, but so much more needs to be done. This research aims to contribute to building a foundation of experimental design that is compatible with the unique nature of machine learning experiments, specifically when applied to machine listening.

References

C. Drummond and N. Japkowicz, “Warning: Statistical benchmarking is addictive. kicking the habit in machine learning,” J. Experimental Theoretical Artificial Intell., vol. 22, pp. 67–80, 2010.
N. Japkowicz and M. Shah, Evaluating Learning Algorithms: A Classification Perspective. New York, NY, USA: Cambridge University Press, 2011.
E. R. Dougherty and L. A. Dalton, “Scientific knowledge is possible with small-sample classification,” EURASIP J. Bioinformatics and Systems Biology, vol. 2013:10, 2013.
K. L. Wagstaff, “Machine learning that matters,” in Proc. Int. Conf. Machine Learning, 2012.

Note: I am interested in supervising other projects related to signal processing and music informatics; please contact me to discuss.

Return to top

Intelligent Machine Listening


Research group: Centre for Digital Music

In this research, we seek ways to exploit novel and holistic approaches to evaluation for building machine listening systems (and constituent parts). A major emphasis will be on answering “how” systems work and “what” they have learned to do, in relation to the success criteria of real-world use cases. The research will involve working at the intersection of digital signal processing, machine learning, and the design and analysis of experiments.

References

K. L. Wagstaff, “Machine learning that matters,” in Proc. Int. Conf. Machine Learning, 2012.
O. Pfungst, Clever Hans (The horse of Mr. Von Osten): A contribution to experimental animal and human psychology. New York: Henry Holt, 1911.
B. L. Sturm, “A simple method to determine if a music information retrieval system is a “horse”,” IEEE Trans. Multimedia, vol. 16, no. 6, pp. 1636–1644, 2014.
B. L. Sturm, “The state of the art ten years after a state of the art: Future research in music information retrieval,” J. New Music Research, vol. 43, no. 2, pp. 147–172, 2014.

Note: I am interested in supervising other projects related to signal processing and music informatics; please contact me to discuss.

Return to top

Automatic music genre recognition


Research group: Centre for Digital Music

As a composer, I am confounded by research that aims to build a computer system that labels a “piece of music” as belonging to one of a handful of speciously defined, culturally negotiated, dynamic, and dare I say, “fetishistic,” “genres" that are supposedly split apart and redefined to suit whatever biases one may have to the tune of, e.g., “I listened to math rock before it was cool.” But enough about me and my “hangups.” This is about you. No – this is about us, versus a world where we can’t deny that music “genres" exist, and people use the vocabulary of genre to discuss, perform and make music, to organise their music collections, to make recommendations, and to discover new music (but apparently not my music). This topic has been researched a considerable amount in music informatics, but almost all of the time “genre” is defined implicitly via some “cleanly labeled” dataset. It is time to take a dramatically different approach, unifying signal processing and machine learning with use case definitions and expert domain knowledge.

References:

O. Pfungst, Clever Hans (The horse of Mr. Von Osten): A contribution to experimental animal and human psychology. New York: Henry Holt, 1911.
J. Frow, Genre. New York, NY, USA: Routledge, 2005.
F. Holt, Genre in popular music. The University of Chicago Press, 2007.
R. O. Gjerdingen and D. Perrott, “Scanning the dial: The rapid recognition of music genres,” J. New Music Research, vol. 37, pp. 93–100, Spring 2008.
B. L. Sturm, “A simple method to determine if a music information retrieval system is a “horse”,” IEEE Trans. Multimedia, vol. 16, no. 6, pp. 1636–1644, 2014.
B. L. Sturm, “The state of the art ten years after a state of the art: Future research in music information retrieval,” J. New Music Research, vol. 43, no. 2, pp. 147–172, 2014.
B. L. Sturm, “A survey of evaluation in music genre recognition,” in Adaptive Multimedia Retrieval: Semantics, Context, and Adaptation (A. Nu ̈rnberger, S. Stober, B. Larsen, and M. Detyniecki, eds.), vol. LNCS 8382, pp. 29–66, Oct. 2014.

Note: I am interested in supervising other projects related to signal processing and music informatics; please contact me to discuss.

Return to top

Evaluating User Experience using Machine Learning


Research group: Centre for Digital Music

Using interactive technologies to create new kinds of collaborative interaction where multiple people create together at once either co-located, or remotely. The project would involve using and developing Machine Learning techniques to systematically evaluate the engagement between people in these creative experiences.

Return to top

Machine-learning architectures for making sense of bird sounds


Research group: Centre for Digital Music

Natural sounds such as birdsong have rich time structure. The "information" in birdsong is not just in the "notes" that are sung, but also in the sequence, the durations, and the size of the pauses between. Yet machine learning systems often use very simple assumptions about temporal structure. How can we build machine learning systems that reflect and represent the rich time structure of birdsong? Possibilities include neural network architectures such as Long Short-Term Memory (LSTM) or Denoising Auto-Encoders, unsupervised feature learning, or hierarchical probabilistic models. How will we know success? Well, possibly by building a better bird recognition system. But further! Maybe the system can explain to us, in words or numbers, what makes a robin sound different from a blackbird. Maybe the system can learn the structure of one species' song, and then go on to sing new melodies it has composed for itself. This topic is suitable for someone with a background in a mathematical topic such as machine learning or statistical optimisation, ideally with a passion for sound.

Return to top

Birdsong similarity measured through structured temporal models


Research group: Centre for Digital Music

Birdsong is highly complex and structured. Yet, even human listeners can often make robust judgments about the relative similarity/difference of two bird song recordings, for example when distinguishing between two different bird species, or two individual birds, or two song types produced by the same individual. Animal communication researchers now have large amounts of data and so they need automatic similarity judgments to help them uncover patterns of cultural and evolutionary transmission. Automatic techniques are commonly used already (cross-correlation, dynamic time warping, Markov models) but these methods are known to ignore complexities such as hierarchical structure, and so their judgment is not rich enough to reproduce perceptual similarity as well as is needed. In this project you will be based in the Machine Listening Lab (in the Centre for Digital Music), using statistical models, machine learning and signal processing to develop a method for similarity measurement that can be applied to a wide variety of bird sounds. Methods might include hierarchical Markov models, state-space models, probabilistic grammars, semi-Markov models, recurrent neural networks (RNNs), information dynamics. You will be collaborating with Dr Rob Lachlan (School of Biological and Chemical Sciences) to work with rich scientific datasets of recorded bird sounds.

Return to top

Deep Learning for Music Compression: Adversarial Training with Attention Modelling


Research group: Centre for Digital Music

The development of music compression was at its peak between 1990-2000. At the time, the automatic identification of perceptually irrelevant information in audio streams enabled a drastic decrease in bitrate and had a massive impact on how we listen to, buy and produce music today. Since then, however, the underlying concepts have hardly changed and can be found in essentially identical form on all subsequent audio coding standards (MPEG surround, SAOC, MPEG-H,...). In lossless audio compression, linear prediction remains a central component that has not been replaced since the 1970s. The goal of this PhD is to design deep neural networks that can learn to anticipate music signals and thus to create a conceptually novel approach for further reducing the bitrate. Such a method can (but not necessarily must) be combined with existing perceptual coding techniques. This problem scenario could involve various types of networks for temporal modelling (possibly with a specific focus on attention-based models to incorporate offline knowledge) and adversarial training (as music streams are compared to the frame-rate of the spectral front-end rather stationary processes and adversarial training helps learning the behaviour at boundaries between such processes).

Return to top

Context-Adaptive Sound Separation with Deep Networks


Research group: Centre for Digital Music

Given an audio recording of a piece of music, the goal of sound separation is to isolate and un-mix individual sounds (related to instruments, notes, melody, background) from the recording. Sound separation is often referred to as a key technology in music processing as it enables direct access to individual sound events and thus increases ways to analyse, process and re-use a recording (e.g. for re-mixing and up-mixing sounds, or for musical analysis). Unfortunately, without making prior assumptions about the recording, sound separation is mathematically ill-posed and unsolvable. However, if rich prior knowledge can be provided, the results are useful for a variety of applications. The goal of this PhD is to investigate in the context of audio sound separation, different approaches to incorporating prior knowledge into deep neural networks or to imposing informed constraints on the learning process. This way, building on recent success in deep learning, the prior knowledge can be used to constrain the expressive and analytical power provided by neural networks, resulting in potentially higher separation quality. In this context, the use and development of (Bayesian) graph networks, variational autoencoders and other methods combining probabilistical modelling with neural networks is of central interest.

Return to top

Deep Learning for Evaluation: Differentiable Metrics from the Crowd


Research group: Centre for Digital Music

Evaluation is of central importance not only in music informatics but in signal processing and machine learning in general - only with proper evaluation it is possible to state whether a proposed system is likely to do what we want it to do. Therefore, the more accurate and representative the evaluation, the more conclusive the results. For example, in music processing, listening tests are often the only reliable way to obtain perceptually relevant evaluation metrics. In machine learning, however, such evaluations have to be conducted thousands or millions of times, as the parameters of a system are adjusted to maximize an evaluation measure - here, listening tests are too time-consuming and expensive. As a makeshift, most machine learning approaches employ simplistic evaluation measures such as a Euclidean distance or a cross-entropy. The goal of this PhD is to develop neural networks that can learn to predict the results of listening tests, with training data obtained using crowd-sourcing services such as Amazon Turk or similar platforms. Such a network is intrinsically differentiable and thus is immediately useful as an objective function in machine learning and thus can directly be used to improve upon the currently used simplistic measures. In this sense, the goal of this PhD is not to replace listening tests but to improve upon how we train our models in general.

Return to top

[SAMI 2] Enhancing the music listening experience


Research group: Centre for Digital Music

In the near future, music formats that allow the listener to build his/her own individual mix from separate instrument tracks will be readily available. Already there is an MPEG standard file format (Interactive Music Application Format - IMAF) that supports this. This project will investigate several of the following ways for listeners to interact with the music and participate more deeply in the listening experience: i) appropriate balance between different instrument stems depending on the acoustic context (e.g. busy subway vs. quiet sitting room); ii) navigation within music pieces according to the musical structure (e.g. skip to chorus, repeat verse without guitar); iii) other active-listener oriented applications, such as intelligent karaoke. Students will need good familiarity with Signal Processing, with User Interaction (or be willing to learn quickly), and with coding in an appropriate programming language. They will work with some MPEG standards, particularly IMAF.

Return to top

[SAMI 1] Studio Science: improving feature extraction in the studio; delivering new experiences to the consumer


Research group: Centre for Digital Music

This project aims to explore in a robust and complete way, the sort of signal processing that should be deployed in the music recording studio in order to extract musical information from instruments: information that can be used to enhance the music listener experience. The project should also therefore understand the needs of music listeners in this changing musical landscape and develop a proof of concept demonstrator to be trialled with users, using the musical information in innovative ways. By ‘Studio Science’ we mean several things. This includes how to extract the best information from music recordings of instruments for: chord estimation, key signature, time signature and rhythm, musical structure and so on. To date although analysis algorithms abound, there are no comprehensive studies to understand how they work on individual instruments, and how to make them work optimally. Which algorithm suits which instrument and with what parameter settings? What resolution is necessary in the time domain to deliver the necessary response time to the consumer wanting to play along with a track? The project requires a good level of signal processing background and an understanding of scientific method. The student will gain experience of High Performance Computing, including GPUs, and will be able to explore new techniques for User Experience.

Return to top

Research group: Centre for Digital Music

Some music services have 30 or 40 million tracks to choose from. So how do you find things to listen to that you don't already know and that will interest you? Well today, it's not easy, and perhaps the best way is to listen to a broadcast radio station that you trust, because the DJ presenter is deliberately searching out "interesting" music. But what if you can do something similar automatically using signal processing? This project aims to do that. Taking inspiration from psychological studies that show that surprise and change in music is interesting, the idea behind the project is to develop new audio features that capture 'interestingness', 'variety' and 'structured-ness' in a piece of music. Starting with pilot studies that determine which features are capable of representing 'interestingness' etc in musical pieces known to be interesting, the project will proceed by working with our partners who aggregate massive music collections (e.g. 40 million songs) and perform user trials. Students will need good familiarity with Signal Processing, with User Interaction (or be willing to learn quickly), and with coding in an appropriate programming language, especially for mobile platforms (or be willing to learn). The student will work with some MPEG standards. One aspect of this project will be to see how much benefit can be found by using the full stereo music signal to increase the discriminatory powers of the features.

Return to top

[SAMI 4] Note level audio features for understanding and visualising musical performance


Research group: Centre for Digital Music

We all know that musical instruments can be played with different levels of skill (or virtuosity). However current Music Informatics research tools rarely address this, largely because the signal processing is not straightforward. This PhD project will investigate performance at the note level so that fine detail can be captured – for example so that vibrato can be measured, parameterised, and maybe used as a retrieval feature. The project will build on prior work from c4dm on high resolution sinusoidal models, and will derive and investigate new ways to visualise and understand musical performance. Once these virtuosic performance aspects are parameterised (and turned into metadata) they can be edited to enhance the performance in subtle ways, such as morphing one performance into another. Imagine pasting a violinist’s vibrato onto a Death Metal vocalist’s singing! The student will need good familiarity with signal processing and programming. There might be some use of Deep Learning and/or Convolutional Neural Nets. One aspect of this project will be to see how much benefit can be found by using the full stereo music signal to increase the discriminatory powers of the features.

Return to top

[SAMI 5] Audio features for MIR based on human hearing physiology and neuroscience and on acoustics


Research group: Centre for Digital Music

In Music Informatics today, researchers tend to choose from a well established set of possible signal processing features to help them extract meaning from the audio. Typically, these features are computed over time windows of 10s of milliseconds to several seconds. This project seeks to expand that toolset by undertaking a thorough investigation of some less used but highly appropriate signal features, both of which are capable of finer resolution in the time domain without sacrificing frequency domain resolution. In order to understand how human hearing physiology can inspire better features, the work will seek to use the instrumental separation implicit in stereo recordings and integrate this into a framework using computational models of human hearing, the best known of which are the so-called gammatone filterbanks. These can provide information on both timbral and harmonic aspects of the music. In order to exploit the acoustic properties of many musical instruments, another aspect of this research is to use Linear Predictive Analysis and Coding (LPC). Here we expect that both the analysis parameters and the residual will capture useful information. One likely outcome is the development on new chromagram-like features (used to represent harmonic content in music) and their use in structural segmentation of musical pieces. Students will require good familiarity with signal processing and programming.

Return to top

[SAMI 6] Compression of individual instrument stems for compact multi-track audio formats


Research group: Centre for Digital Music

Music formats are evolving. Before long, listeners will buy or stream music that contains all the individual instrument tracks, as well as other interesting information. A 64-track master recording at studio-grade audio quality contains many times as much raw data as an MP3 or AAC version of the same piece. So this project aims to address this data explosion by looking at ways to compress the multitrack music version. As a start, each instrument could be separately compressed using specific algorithms that account for the instrument's unique nature (e.g. that it is an acoustic resonant system implying the use of LPC, or that it uses additive synthesis implying the use of sinusoidal modelling) and that silence can be effectively encoded. Subsequent coding gains can be obtained by exploiting correlations between instruments as well as using any musical score or lead sheets. Students will need good familiarity with Signal Processing and with coding in an appropriate programming language. Interest in Deep Learning and Neural Networks is advantageous. They will with some MPEG standards, particularly the IMAF (Interactive Music Application Format) standard for multi-track audio.

Return to top

Mutually Engaging Interaction


Research group: Centre for Digital Music

Using interactive technologies to create new kinds of musical interaction where multiple people create music at once either co-located, or remotely. The Interaction Design would involve systematic evaluation of the engagement between people in these creative experiences. Similar research includes Daisyphone research: http://isam.eecs.qmul.ac.uk/projects/daisyphone/daisyphone.html

Return to top

Cross-Modal Interaction


Research group: Centre for Digital Music

Interactive technologies give us the chance to design engaging experiences for people with different sensory abilities. This project would investigate how to design effective, intuitive, and engaging cross-modal interaction. The focus of the Interaction Design would be on how to evaluate the experiences of cross-modal interaction. Similar research includes: http://isam.eecs.qmul.ac.uk/projects/ccmi/ccmi.html http://isam.eecs.qmul.ac.uk/projects/depic/depic.html

Return to top

Wearable Interaction Design with Audio


Research group: Centre for Digital Music

This project would involve building and evaluating prototype wearable interfaces e.g. using accelerometers as input, focussing particularly on using sound in innovative ways to convey information to users. Evaluation would focus on understanding the user experience in a range of domains such as interactive art, performance, or day-to-day interaction through sound. http://isam.eecs.qmul.ac.uk/projects/sensory_threads/sensory_threads.html http://isam.eecs.qmul.ac.uk/projects/whimsichord/whimsichord.html

Return to top

Interaction Design with Audio


Research group: Centre for Digital Music

Undertaking Interaction Design of novel interactive experiences focussing on the role of Audio in interaction. Would involve some build and making. Some user centred testing and evaluation. Could include tangible or virtual interaction forms. Example projects are are: http://isam.eecs.qmul.ac.uk

Return to top

Modelling the auditory brain as a predictive machine


Research group: Cognitive Science

Theoretical models propose that attention and prediction are separable, independent processes but research in visual neuroscience suggests that the predictability of the stimulus may also guide attention. Recent research has shown that listeners are capable of very quickly detecting the emergence of regularities in timing and frequency patterning in auditory scenes. Predictions about the timing or frequency of auditory events are thought to reflect a hierarchical process of top-down prediction at a range of different time-scales based on a model of the sensory environment. The model is learned adaptively through exposure and the discrepancy between predicted and actual sensory input (the prediction error) is used to drive learning. Predictions are thought to be precision-weighted such that very certain (low entropy) predictions gain a higher weight than less certain (higher entropy) predictions. Recent research has revealed increased power in neural responses to more predictable scenes which may reflect this process of increased gain for low entropy predictions. The proposed project will investigate the relationship between attention and prediction using experimental manipulations that orthogonalise these two top-down processes by varying predictability within components of the auditory scene while manipulating attention between different components. The project will also involve studies that systematically vary the degree of predictability and strength of attentional manipulation. This work will involve both psychophysics and EEG and will be complemented by computational research to develop probabilistic models that simulate the interaction between prediction and attention in perception of auditory scenes. There is scope to investigate both pitch regularities and temporal regularities and combinations of the two.

References

Barascud, N., Pearce, M. T., Griffiths, T. D., Friston, K. J., & Chait, M. (2016). Brain responses in humans reveal ideal observer-like sensitivity to complex acoustic patterns. Proceedings of the National Academy of Sciences, 113, E616-E625.

Pearce, M. T. (2005). The Construction and Evaluation of Statistical Models of Melodic Structure in Music Perception and Composition. Doctoral Dissertation, Department of Computing, City University, London, UK.

Return to top

Quantum Methods for Analyzing Music and Language as Communicative Systems


Research group: Cognitive Science

Supervised by Mehrnoosh Sadrzadeh and Marcus Pearce.

Quantum probabilities have been used in language n-gram models to reason about meaning representations of word compounds and their performance has been tested versus the bag of words model; they have also been used to model cognitive concepts and how they are represented in the human brain. Notably, quantum methods have been successfully applied to understanding representation of meaning and, in particular, ambiguous concepts. The purpose of the project is to apply these methods to music. Music is also a communicative system that consists of sequences of auditory symbols conveying meaning, which may be ambiguous. Both language and music have been successfully modelled using probabilistic methods. Ambiguity in music arises with respect to structural interpretations of key, metrical interpretation, grouping structure and emotional interpretation. The project will use quantum methods to create computational models that are capable of representing and potentially resolving ambiguity in these aspects of music. The models will be evaluated in terms of how well they simulate human music perception making a contribution to the new field of quantum cognition.

Return to top

Developing a privacy-preserving Mobile/IoT sensing system


Research group: Cognitive Science

This studentship will be part of the Databox project http://www.databoxproject.uk where we investigate and build a privacy-preserving app ecosystem. In this studentship we will focus on aggregating and analysing data from smartphones and home IoT devices, while performing analytics at the user/device end to save on bandwidth, energy, and privacy costs.

Return to top

Distributional Pragmatics


Research group: Cognitive Science

To get computers to understand conversational data, we need models of meaning in dialogue. Conventional methods depend on knowing some taxonomy of dialogue act types, and these are hard to define and quite task- and domain-specific. Recent work in distributional semantics and neural networks has shown that we can directly learn semantics (meaning of words and sentences). Can we do the same for pragmatic aspects of meaning in context, and thus build systems which can learn to interpret dialogue?

Return to top

Probabilistic Modelling of Music Perception


Research group: Cognitive Science

The goal of this project is to understand how listeners perceive and predict musical structure during listening. Expectation in musical listening has an impact on structural interpretation of musical syntax and also the dynamic emotional experience of the listener as the music unfolds in time. Our existing research suggests that listeners' melodic expectations are acquired through unsupervised statistical learning and reflect probabilistic predictions about forthcoming musical events given the preceding context. The goal of this research project is to extend and generalise these findings to rhythmic and harmonic structure in polyphonic music using a combination of computational modelling, behavioural testing of listeners and examination of brain responses using Electroencephalography (EEG).

Return to top

The anatomy of scientific research


Research group: Cognitive Science

This project aim to develop a novel complexity science approach to better understand pattern of collaborations. Traditionally, academic impact is often illustrated by IP generation and bibliometric measures, while economic and societal impacts are extremely diverse and difficult to quantify. In particular, research is largely not carried out in isolation but through collaborative activities, and therefore, impact should also be understood in a collective and community sense. Collaboration culture and practice found in research communities are highly dynamic as they self organise to adapt to the different levels of research policy, available funding schemes and disciplinary practice. If one can consider a positive impact as better opportunities for knowledge transfer or closer collaborations, both within a community or across different ones, the key challenge is how do we measure it? An example of this kind of impact is the creation of social capital, in the form of structural holes and network closure, through the addition of new collaborations and removal of expired ones; resulting in different forms of evolving local network structures when expressing these collaborations as an innovation ecosystem. This leads to questions such as ’How is social capital developed within an ecosystem?’; ’How are they distributed across an ecosystem?’; ’What are their characteristics and impacts on knowledge propagation?’.

Return to top

Using network theory to evaluate the impact of climate change


Research group: Cognitive Science

Climate change is projected to increase the likelihood of extreme environmental conditions, causing major transformation of ecosystems. These include the loss of species and altered biomass fluxes through the food web. The patterning and strength of species interactions within the food web can play a huge role in determining how the system responds to environmental perturbations, but surprisingly, many food web metrics appear to be insensitive to these major perturbations. For instance, drought is said to triggered partial collapse of food webs in which both species and links are loss and yet, macroscopic network measures, such as connectance and interaction diversity, were unaltered by this disturbance. The project aims to develop a more functional understanding of how higher-level (e.g. food web) properties respond to environmental conditions and stressors. By coupling techniques in network science with ecological theories, we assess a novel aspect of functional biodiversity (species interactions) that has been overlooked in biomonitoring, due to the bias towards focusing on the nodes (i.e. species) within food webs rather than the links (e.g. trophic interactions between species pairs). Network properties, such as nestedness and modularity, have been linked to propagation of disturbance in food webs and impact on network stability, and we are still at an embryonic stage in understanding the functional role of different network structures and properties and their ecological implications. In particular, current assessment on network robustness relies heavily on simulated removal of a large proportion of species, and this approach is, somewhat, unrealistic and it is not effective in capture immediate and short-term impacts of the extinction of a given species. By employing network theory and dynamical modelling, the project aims to gain a better understanding on the effect of perturbations in food webs, so as to advance the predication of stability in complex natural systems and reveal the key underlying organisational properties that contribute towards stability.

Return to top

Databox: Privacy-Aware Personal Data Platform


Research group: Cognitive Science

The Databox envisions an open-source personal networked device, augmented by cloud-hosted services, that collates, curates, and mediates access to an individual’s personal data by verified and audited third party applications and services. The Databox will form the heart of an individual’s personal data processing ecosystem, providing a platform for managing secure access to data and enabling authorised third parties to provide the owner with authenticated services, including services that may be accessed while roaming outside the home environment. The studentship will focus on machine learning applications, social media analysis, and IoT cyber-physical systems to process household data. Databox

Return to top

Mobile Sensing for Crowd Interaction Analysis


Research group: Cognitive Science

Smartphones are an inseparable part of our lives. The range of advanced functionalities and accurate sensors makes them a great platform for performing user studies in a number of areas including health and cybersecurity. In this studentship we will focus on the use and development of www.sensingkit.org as a platform for engagement in a number of personal analytics and monitoring and inter-personal interaction and mobility analytics within a crowd, in collaboration with colleagues engaged in http://researchstack.org/

Return to top

Digital Technology and Transparency


Research group: Cognitive Science

Motivation In the aftermath of the 2008 global financial crisis the notion of transparency became a rhetorical token used to provide solutions to the financial problems. There seems to be a contrast, however, between a political understanding of transparency which aims at fairness and stability in the financial system and a technological understanding of transparency, aiming at the efficiency and effectiveness of its (the financial system’s) information infrastructures. Indeed, while political transparency calls for greater visibility and public availability of data on financial transactions, a technological understanding of transparency is rather based on the “invisibility” of the information infrastructures on which these transactions run on. As Star & Ruhleder (1996) have famously argued: “Infrastructure is transparent to use in the sense that it does not have to be reinvented or assembled for each task, but invisibly supports those tasks” (emphasis added). The tensions between the political calls for transparency as visibility on the one hand and the need for transparent infrastructures that invisibly support operational practices, on the other, raise questions in relation to the development, implementation and use of digital technologies, not only in finance but in various organizational and institutional settings (healthcare, government, social media, urban planning etc). Research objectives The questions we ask is: Can digital technologies become effective and efficient transparent infrastructures in the hands of their users, while at the same time enable a visibility that would perform accountability relationships in a fair and stable manner? How can this be achieved in different contexts? What is the role of designers, users and other stakeholders? This topic can be investigated in a variety of contexts. The most obvious is financial markets, however, I would accept proposals in other areas as well (healthcare, social media, government organisations, and so on). Additionally, the digital technologies considered may vary depending on the context. Research approach and contribution Based on an interpretive paradigm it is expected that the candidate should carry out an ethnographic investigation (interviews, observation, etc) of a relevant empirical field. However, they should explore innovative methodological approaches. Findings and analysis should contribute to an interdisciplinary understanding of the relationship between digital technology and transparency by combining insights from various disciplinary areas (Information Systems, Management, Science & Technology Studies, Innovation Studies). Implications for development, implementation and use practice should be drawn.

Return to top

Social media participation and value creation


Research group: Cognitive Science

This project is looking to explore social media participation and value creation. Social media (social networking sites, blogs, content communities etc) constitute the current “face” of the Internet. Studies on social media participation and value creation appear quite fragmented. On the one hand, user-centric studies which emphasise agency, participation and sense-making understand value creation in the context of an exploration of the self and management of social relations of the users; on the other hand, industry-centric studies focus on engagement and profiling understand value in economic terms (Bechmann & Lomborg, 2012). In the former, social media users appear as empowered agents in pursuit of their identities, while in the latter, they are treated as a resource by companies (i.e. for market research, for viral marketing campaigns etc). This project is set out to provide a bridge between the two approaches. In doing so, the perspectives of social media users shall be explored and also the points of view of industrial organisations. We ask, is there evidence of co-creation of value between social media users and corporate organisations? Is it always conscious and ethical? What is the role of technology in mediating and performing value relations? Methodologically, the candidate should consider a mixed method approach which combines traditional qualitative and quantitative techniques, with emerging digital methods (i.e. digital ethnography/netnography). Contributions should be multidisciplinary and combine insights from new media studies, sociology, information systems, science and technology studies.

Return to top

Are we creatures of habit?


Research group: Cognitive Science

The big data phenomena has generated information in an unprecedented rate in terms of its volume, variety and velocity, providing opportunities to curate meaningful information that can be used to address local problems and questions at larger spatial and temporal scales. Online systems (or applications) often have the ability to capture our behaviour with a specific focus; for example, social networking applications such as Facebook provides information on friendships and interactions among friends; however, in reality, our cyber presence is richer and much more heterogeneous due to the vast number of application domains, and we are still at an embryonic stage in understanding the underlying patterns. Interestingly, Information from one context can be useful to make sense in another context. For example, links in one social network (Facebook) can be indicative of interactions in another social network (Pinterest/last.fm), images in Pinterest could link to sales in ebay, and opinions on twitter could have an influence on brand perception. Here, we aim to develop novel techniques to identify interrelated patterns from data across different systems and examine the underlying dependency, which will help in devising strategies in converting data into useful information using a reduced number of data sources but yet providing information that will be suitable for cross-platform exploitation. In addition, knowledge on the co-variance between systems will provide an insight into areas that are unique to their respective systems, and information of this kind can be used to address local issues or enhance local decision.

Return to top

Interaction Mining


Research group: Cognitive Science

With the large datasets of human interaction we have available today, we often want to find and summarise specific meaningful events: decisions made in business meetings, answers to questions in radio broadcasts, agreements and disagreements in social media. This is a challenging problem: rather than identifying keywords or specific sentence structures categories, we must identify structural patterns in the interaction itself (e.g. the characteristic conversational patterns of the decision-making process). The events in question are also often rare, and often involve more than two people: structures are consequently more complex, less predictable, and less well suited to common dialogue models. This project will develop methods which combine rule-based models of dialogue structure with robust statistical approaches to solve this problem, exploiting recent advances in NLP and machine learning such as distributional models of meaning and deep learning.

Return to top

Language Processing for Mental Health Treatment


Research group: Cognitive Science

Recent work within EECS and elsewhere has shown that natural language processing and machine learning can help us characterise mental health conditions by analysing patterns in the language people use, with progress in diagnosing conditions and symptoms and in predicting the outcomes of therapy. This initial success opens up a new field of research with several challenges. Can computational models help us understand mental health treatment, by providing empirical models of the processes within therapy and their association with outcomes? Can we improve clinical systems by helping provide earlier or more accurate diagnosis, particularly with dementia? This project will build on recent work in the Cognitive Science group in collaboration with researchers in QMUL's School of Medicine and the universities of Exeter and Warwick, to advance both computational linguistic techniques for analysing and understanding dialogue, and their link to clinical research.

Return to top

Computer based histopathological image analysis for cancer grading and progression assessment


Research group: Multimedia and Vision

The assessment of pathological cancer regression after preoperative chemotherapy is mostly based on the assessment of tumour morphological features, such as the proportion of cancer cells in relation to the total tumour region, as well as biologically relevant histology features, such the tumour invasion front. Currently, this histopathological evaluation is performed by expert pathologists through visual assessment of the tumour microscopic slides. This is often time-consuming, expensive and may be unacceptably inconsistent and imprecise. This project aims at developing an intelligent system that enables automatic, precise, objective and reproducible assessment of tumour regression and precise characterisation of the tumour invasion front based on the digital scans of resected tumour tissue slides, by integrating beyond the state-of-the-art, specifically designed computer vision, image processing and machine learning schemes.

Return to top

High Quality Semi-Automatic Production of Animations Using Commodity RGB-D Sensors


Research group: Multimedia and Vision

Motion Capture (MoCap) has been used for decades in animation studios to record human motion and replicate or apply such movements to any character in animated films or games. Famous movies as Avatar exploited MoCap technology in many of its scenes. Due to the relatively high cost of hardware and software to accurately track human motion, till very recently this technology has been confined to professional studios. With the arrival of inexpensive (commodity) RGB-D cameras/sensors as the Kinect II a new frontier in more efficient, accurate and affordable animation creation has been opened. This PhD project aims at building on the outcomes of the REVERIE project (http://www.reveriefp7.eu/) which was (technically) coordinated by the MMV Group at QMUL. The aim is to develop more accurate, more stable and more efficient algorithms for character puppeting and animation using autonomously calibrated multi-Kinect II sensors or similar commodity hardware. Sought candidates should be able to show talent in combining artistic design of virtual characters with computer vision and graphics techniques and programming skills for the development of improved software tools for MoCap and highly efficient and realistic character animation.

Return to top

High Efficiency Video Coding


Research group: Multimedia and Vision

High and ultrahigh resolution cameras and displays are becoming pervasive. They enable a more immersive and compelling experience when watching TV, films or any other multimedia content over conventional broadcasting channels or the internet. The format to deliver this improved quality of experience is called Ultra High Definition Television (UHDTV). UHDTV requires 3840x2160 pixels/frame or 7680x4320 pixels/frame and a temporal resolutions of up to 120 frames per second. Clearly, the volume of data associated with UHDTV signals is enormous. Hence, to be able to deliver such signals over conventional networks, efficient video compression technology is needed. In response to this acute need, the ITU Video Coding Experts Group (VCEG) and ISO Moving Picture Experts Group (MPEG) joined efforts in a partnership called Joint Collaborative Team on Video Coding (JCT-VC) to develop the High Efficiency Video Coding (HEVC) standard. Version 1 of HEVC was finalised in January 2013 and proved to outperform its predecessor Advanced Video Coding (AVC) by providing up to 50% bitrate reduction for the same perceived quality. For UHD content, the reduction is even higher than 50%. Given the superior performance, HEVC will be the best candidate in the deployment of UHDTV services. As expected, this improved compression efficiency comes at the expense of increased complexity. Therefore, any practical implementation must optimise HEVC coding by reducing the complexity without sacrificing the compression performance. This PhD project will develop techniques to greatly reduce the complicity of the HEVC video coding standard. The aim is to provide tools for HEVC that can be implemented or ported to commercial video coders building on the related comprehensive expertise of the MMV group at QMUL and previously implemented commercial HEVC coders as the Turing codec developed in cooperation with The BBC (http://turingcodec.org/)

Return to top

Mosaics and Demosaicking of Colour Filter Arrays


Research group: Multimedia and Vision

Almost every modern camera has a colour filter array (CFA) fabricated on top of the light sensors for creation of colour images. The CFA-filtered images acquire energy of only one colour light at each pixel and a technique known as “demosaicking” follows to reconstruct the images with all the three primary colours, red, green and blue, at each pixel. The final image quality depends on the mosaic of the CFA and the demosaicking method. Having worked on such a subject for a few years, we have achieved a methodology for optimal regular CFA design and a universal demosaicking method. Based on these, we wish to develop further theory and techniques for better imaging quality, to find optimal irregular CFAs with constraints of some practical requirements and to build the associated best-fit demosaicking algorithms which should be fast and memory-efficient and give the high image quality. With the new achievements on mosaics and demosaicking of colour filter arrays, research would be done to develop models of human vision system, specifically the low-level vision. This may lead to better understanding of human vision.

Return to top

Best prediction of perceived user quality in mobile networks


Research group: Networks

Enhancing the traditional network QoS metrics of packet loss, delay and jitter, mobile network operators increasingly rely on a new metric called “QoE” – Quality of Experience. QoE provides an objective measure of the end user’s satisfaction level. For example, Huawei has recently introduced U-vMOS as a user experience grading system for videos, (http://www.huawei.com/en/news/2016/4/U-vMOS-Supports-for-Video-Communication). As well as international organisations like Huawei, many smaller companies have recently started up in this area. These new companies offer specialist QoE measurement services, e.g. https://www.thousandeyes.com/ and https://www.actual-experience.com/. QoE is often predicted through measurement of the QoS metrics of packet loss, delay and jitter. However, currently there are no studies of how accurately and precisely QoS sampling enables the prediction of QoE. Preliminary work at QMUL has shown that, unless operators can account for the errors in precision and accuracy, they may predict GOOD QoE when the QoE is POOR, and POOR when it is actually GOOD. Because of its growing importance such errors in QoE prediction can now have huge commercial implications. The aim of this research project is to discover the fundamental networking relationships that affect QoE prediction, and ultimately to guide network operators in making the best use of their QoS measurements so that they can optimally predict end user QoE.

Return to top

Routing based Security Enhancement scheme for MTC in E-health Network


Research group: Networks

NHS has allocated £4.2bn of additional transformation funding for technology programmes from 2017/18 to 2020/21. It intends to use digital approaches to support new models of care in general practice and to support daily submission of electronic Secondary User Service (SUS) data from April 2018. The overall 2020 goals include paper-free at the point of care, 95% e-consultation and 95% digitally transferred tests between organisations. This project focus on security issue in Machine Type Communication (MTC) in E-health monitoring system. Under the network structure, various types of wireless devices, such as sensors, wearable devices and smartphones, can automatically conduct the communication for collecting, transmitting and processing health related data in a self-organized matter. It is believed that massive small size data will be generated by various electronic devices and transmitted through MTC network every day. However such network is vulnerable to various security attacks, from both inside and outside MTC network. The idea in this project intends to develop a routing based scheme to introduce distributed security mechanism in Network layer to enhance the security in MTC network for providing e-care monitoring services. This project intends to benefit potential users, such as aging populations, to provide flexible, stable and long-lasting e-care monitoring aid. The work will be primarily verified and evaluated through simulations; however hardware related verification will also be considered, time permitting.

Return to top

Understanding Mobile Social Apps


Research group: Networks

Mobile apps are becoming an increasingly prominent part of our daily lives (e.g. Uber, City Mapper). Social apps are those that entail some forms of social interaction with those around you. These include messenger services (WhatsApp, Snapchat), social networks (Facebook), micro-blogging (Twitter) and games (Pokemon Go). A recent brand of these even include the capacity to discover and interact with people nearby (e.g. Tinder). This PhD project will involve exploring the use of these social apps, to gain an understanding of how people interact with them (and each other). This will, of course, require the collection and analysis of mobile user data. It may also require the development of mobile apps that can be deployed. The goal of this project is to shed light on the (current) poorly understood nature of these mobile social apps.

Return to top

Exploring the fundamental nature of video traffic


Research group: Networks

The vast bulk of traffic on the Internet today is video traffic. To understand the nature of the Internet today it is therefore vital that researchers gain a better understanding of the nature of online video. Most video traffic has a specific nature in that it is very "bursty" because of the way that providers choose to send traffic. It often appears as short spikes with a large amount of traffic and then quiet periods of several seconds with no traffic. There are important open questions about what happens when several users in the same network (for example in the same house) are watching video content at the same time. If the bursts happen together this could create packet loss and hence worse playback. Understanding and preventing the problem could greatly improve the experience of watching video on shared networks. This project will involve practical measurements and analysis of real traffic both at end user networks and also at intermediate points in the network. The goal of the project is to answer several important research questions: Can we develop a unified model of video traffic that estimates its character at all points on the network and that explains the real measured data? Can we show how multiple video sources combine on a home network? Can we use this knowledge to improve user experience when several users are viewing different video on the same network? This project will be jointly supervised by Dr Gareth Tyson and Dr Richard Clegg

Return to top

Exploring the African Internet Ecosystem


Research group: Networks

Africa has the fastest economic growth of any continent today. In line with this, rapid investments are being made in the regional Internet infrastructure. This means that it is evolving and expanding in unique ways, very different from the infrastructure seen in Europe. This PhD project will involve trying to measure and analyse how this infrastructure is changing, as well as how it differs from the infrastructure in other continents. This is of critical importance to help inform its future development. In this project, it will be necessary to design, develop and deploy "measurement software" across the world to collect meaningful data regarding Africa's network paths, server locations, websites, mobile technologies etc. Through this, the student will gain an expert understanding in how the Internet in Africa is deployed an operated. Using this expertise, the project will then progress to propose technologies and techniques to improve any issues found.

Return to top

Understanding Mobile Social Apps


Research group: Networks

Mobile apps are becoming an increasingly prominent part of our daily lives (e.g. Uber, City Mapper). Social apps are those that entail some forms of social interaction with those around you. These include messenger services (WhatsApp, Snapchat), social networks (Facebook), micro-blogging (Twitter) and games (Pokemon Go). A recent brand of these even include the capacity to discover and interact with people nearby (e.g. Tinder). This PhD project will involve exploring the use of these social apps, to gain an understanding of how people interact with them (and each other). This will, of course, require the collection and analysis of mobile user data. It may also require the development of mobile apps that can be deployed. The goal of this project is to shed light on the (current) poorly understood nature of these mobile social apps.

Return to top

Exploring the fundamental nature of video traffic


Research group: Networks

The vast bulk of traffic on the Internet today is video traffic. To understand the nature of the Internet today it is therefore vital that researchers gain a better understanding of the nature of online video. Most video traffic has a specific nature in that it is very "bursty" because of the way that providers choose to send traffic. It often appears as short spikes with a large amount of traffic and then quiet periods of several seconds with no traffic. There are important open questions about what happens when several users in the same network (for example in the same house) are watching video content at the same time. If the bursts happen together this could create packet loss and hence worse playback. Understanding and preventing the problem could greatly improve the experience of watching video on shared networks. This project will involve practical measurements and analysis of real traffic both at end user networks and also at intermediate points in the network. The goal of the project is to answer several important research questions: Can we develop a unified model of video traffic that estimates its character at all points on the network and that explains the real measured data? Can we show how multiple video sources combine on a home network? Can we use this knowledge to improve user experience when several users are viewing different video on the same network? This project will be jointly supervised by Dr Gareth Tyson and Dr Richard Clegg

Return to top

5G HetNet in the Unlicensed Band: Coexistence between WiFi and Small Cell


Research group: Networks

WiFi/WLAN has been the most dominant technology in the unlicensed band. LTE-U is an emerging cellular wireless technology aimed at providing carrier grade wireless service in the unlicensed band. The standard is particularly targeting 5GHz band where there is more than 400 MHz of bandwidth is available. WiFi is already operating in this band. So, Cellular mobile technology needs to make sure that the impact on the WiFi transmission remain in the tolerable level. Since these two technologies work differently, co-existence of them are not straightforward and a number of challenges need to be addressed before deploying LTE in the unlicensed band. This work will focus on designing co-existence mechanism between WiFi and LTE system with particular focus on 5GHz band.

Return to top

Exploiting D2D communication for energy saving in 5G heterogeneous cellular networks


Research group: Networks

Massive deployment of small cells over the macrocell coverage has been considered as one of the most promising pathways for achieving the vision for future 5G cellular networks supporting very high data rate for users regardless of the location. When a high number of access points are available, at low traffic conditions, a significant number of the AP may also be very lightly loaded which leads to the energy inefficiency. A considerable amount of energy can be saved if the unused small cells APs are switched off or go in hibernation when there is no traffic to particular cells. A further saving can be made if the users of the lightly loaded cells are offloaded to suitable neighbouring cells when available while ensuring the required QoS. However, the amount of user offloading using these techniques are limited as there are constraints over transmit power and coverage range expansion. If this user association techniques can be considered with D2D enabled system, then more number of users can be offloaded since a user can be associated to a cell, even though it is not covered by that access point, when a UE within the coverage of the AP acts as a relay for that user. This project will aim at designing mechanism for exploiting D2D communications for freeing some lightly loaded cells and switching them off to ensure further energy saving.

Return to top

Mountain Rescue Communications Support


Research group: Networks

The aim of this project is to provide communications support to mountain rescue teams. In remote areas of mountains there are regions where there is no GPS, no GSM and no electricity but there is, in an emergency, a need for people to communicate to the mountain rescue volunteers or police. It is assumed that people will carry a mobile device such as a smart phone. Hence the technology used must be able to talk to such a device. This project is looking at developing a robust, reliable and low energy communication system that will allow people to use their mobile devices to communicate with the mountain rescue teams under severe energy/communication constraints. The novel propositions will: 1. Support a localization network able to work in dead zone areas and to reach a GSM/4G/LTE communication point, where there is a signal to communicate to the mountain rescue teams; 2. Support low energy but long range (e.g. 1 km) communication networks run in the critical areas; 3 support communication by WiFi and/or e.g. XRF to and from the GSM communications point from peoples’ smart phones; 4. to use people presence detectors/sensors, e.g. WiFi presence and personal alarm signal to trigger running the higher energy devices; 5.to use historical data and artificial intelligence techniques to help manage the energy usage to maximise the potential to detect people in trouble. It is unlikely that all the devices can be on all the time and hence they should be only on when they are needed or there is a reasonable probability that they would be needed. Different approaches need to be evaluated. The student will be able to modify the approach (with agreement) in directions that make the approach more effective and meet the mountain rescue requirements better. Initial components have been developed but the system needs to be further developed and tested and validated in cooperation with the mountain rescue people. System integration and validation are key aspects of this project.

Return to top

Lightweight and Dynamic Messaging System


Research group: Networks

With the massive growth of the Internet and advances of its applications, many enterprise systems are becoming systems of systems (SoS), as a collection of subsystems distributed over geographically different areas. Financial services, telecommunication applications, transportation and health-care systems can generate a complex set of producers and consumers of information. Another example of the increasing scale of producers and consumers of information is the Internet of Things (IoT), that will be applied to a variety of domains raging from the home (i.e., smart homes), whole countries (i.e., smart grid) to worldwide (i.e., environmental monitoring). Therefore, a very large scale of devices producing/consuming data can be forecasted. In both SoS and IoT become very complex for producers to control all unit demands of information. So information aggregators play a key role to alleviate the high information demand. A publish/subscribe based messaging/content framework is very suitable in providing these services, however effective routing and resilience of delivery are of extreme importance to guarantee correct and efficient delivery of information. This project proposes a light content driven routing for message/content delivery systems, which can allow profile generation, policy compliance and replacement ability in emergency and warning scenarios. The project will investigate a lightweight description and matching techniques for routing and delivery, but the major innovation will be that the middleware complexity and behaviour of description changes with the environment and communication current context, allowing new advertisements from the ground to be incorporated and to change the content of the messaging system. This is a novel approach for use in overlay/domain networks and for the IoT data aggregation and delivery. The student will extensively review currently proposed solutions in the literature and research feasible modifications leading to a novel proposal(s) for the scenarios envisaged. IBM is interested in this type of research and the supervisor will propose an internship for the student. Depending of the policies proposed in the research, BT may also be interested. The prospect supervisor recently had one PhD student in an internship in BT.

Return to top

SDN-based Energy-aware Management for Campus Networks


Research group: Networks

Given that the vast majority of Internet energy consumption is associated with access networks, the focus of this research is the “green campus”. In a campus environment, a diverse portfolio of equipment is owned by a single organisation. The goal is to explore the extent to which coordinated cross-layer mechanisms can reduce energy usage whilst ensuring the organisation itself is the key beneficiary. Possible mechanisms to save energy include adaptive packet-scheduling, the use of rate adaptation and device mode-switching such as Sleep-On-Idle, with some wireless access points changing to an adhoc relay mode to enable them to switch off their wired network interface cards. Building on previous research, we will examine the trade-off between responsiveness, energy consumption and QoE impact within the organisation. Our direct links with Juniper Networks will enable the use of the Junosphere cloud environment and its supporting SDK to evaluate the performance of proposed mechanisms. It is expected this research will lead to the creation of energy-efficient next generation devices and architectures incorporating feasible state-of-the-art mechanisms.

Return to top

Resource Allocation for NOMA-based D2D Communications


Research group: Networks

Device-to-device (D2D) communications is considered as one of the pieces of the fifth generation (5G) jigsaw puzzle in order to improve spectral efficiency. Apart from invoking D2D technique to improve the spectral efficiency of the wireless networks, another emerging technique, non-orthogonal multiple access (NOMA) is able to address the spectral efficiency enhancement issue, on the standpoint of realising a new power dimension for multiple access. Similar to D2D communications, Non-orthogonal multiple access (NOMA) has also been recognised as a promising technique for 5G due to its high spectral efficiency and user fairness. Different from the conventional orthogonal multiple access (OMA) technique, NOMA is capable of supporting users to share the same resource (e.g., time/frequency/code) by performing successive interference cancellation (SIC) with different channel state information (CSI). However, perfect CSI is difficult to obtain in practice due to either the estimation error or the feedback delay. This work research into how CSI error affects the performance of NOM-based D2D communications and propose a robust resource allocation under imperfect CSI.

Return to top

Internet of Things: Friend or Foe


Research group: Networks

The industrial an academic supporters of the Internet of Things (IoT) promote its deployment as they argue it would take us from the Information Age to the "Intelligence Age". The opponents argue that the IoT would bring a major loss in our privacy and the possibility of catastrophic failures as it would create strong interdependencies, for example, the food supply could depend on GPS for its distribution but also in RFID tags for animal welfare and, these two seemingly disparate sources of information would be correlated via the implementation of the IoT. This research is about how to describe interdependencies in large communication network, being physical (router) or logical (software) and if these interdependencies can be used to predict a catastrophic failure.

Return to top

Enhanced Security for Multimedia Big-Data Transmissions in CRMANET


Research group: Networks

Cognitive Radio (CR) enabled Mobile Ad hoc Networks (CRMANET) are a rich research area nowadays owing to their infrastructure-free characteristics, which means that they are more flexible than traditional wired networks and comprise powerful mobile devices with various sensing capabilities. However, the lack of a centralized authority together with device-to-device communication characteristics can make them more vulnerable to malicious attacks. Recently, the fast growing multimedia based big data communication in wireless networks has brought a lot of concerns for security. Many researchers have dedicated their work to find the solutions from the aspects of data encryption/decryption, channel coding/decoding, cipher key generation/protection, etc. however, such actions require a great deal of computational effort on each involved node in the ad hoc networks. This project aims to find a solution from MAC/Network Layers to provide secured route for data transmission, distributed data compression and security enhancement are also desired in the scheme. Meanwhile a system level efficiency need to be guaranteed and energy consumptions should be maintained as minimum level without harm the quality of service. The overall intention is to pro-actively protect against potential threats while multimedia data is transmitted and compressed in a distributed matter in CRMANET and make such networks more secure and resilient during the data transmission process. The work will be primarily verified and evaluated through simulations.

Return to top

Edge caching in Small Cell and D2D Networks


Research group: Networks

Given the exponential growth of mobile data traffic and the consistent decline of the cost per stored bit due to the fast development of capacity of modern storage units, edge caching has drawn increasing attention from industry and academia as a technology which is able to alleviate the burden on backhaul links in small cell networks and at the same time increase quality of experience (QoE) by reducing content delivery delay. In edge caching, popular content can be cached at the network edge, either at a small cell base station or directly at the user terminals for Device-to-Device (D2D) communications. There are many challenges in designing efficient edge caching strategy in mobile networks due to various constraints and trade-offs, which need to be taken into consideration. For example, constraints on energy and bandwidth, and trade-off between density of small cells and storage capacities. We are offering a PhD candidate the opportunity to investigate and develop frameworks for modelling and evaluating the efficiency of edge caching strategies for user experience enhancement in small cell and D2D networks.

Return to top

Resource Allocation Optimization for mmWave Fused Heterogeneous Networks


Research group: Networks

mmWave communication with the new spectrum harnessed at higher carrier frequencies constitutes one of the most salient enabling technologies envisioned for future fifth generation (5G) networks. 5G networks are expected to become the fusion of heterogeneous wireless technologies, mmWave networks will coexist with various networks, such as the Wi-Fi and legacy cellular networks, so as to achieve the full potential of 5G networks. These intricate network conditions pose substantial challenges to the radio resource allocation optimization therein. We are offering a PhD candidate the opportunity to tackle the radio resource allocation problem and specifically target the mmWave fused heterogeneous networks scenario, in order to satisfy the increasing demand for wireless broadband access and a wider range of requirements for 5G and beyond, including higher capacity, higher date rate, reduced latency, improved energy efficiency and massive device connectivity.

Return to top

Exploring the African Internet Ecosystem


Research group: Networks

Africa has the fastest economic growth of any continent today. In line with this, rapid investments are being made in the regional Internet infrastructure. This means that it is evolving and expanding in unique ways, very different from the infrastructure seen in Europe. This PhD project will involve trying to measure and analyse how this infrastructure is changing, as well as how it differs from the infrastructure in other continents. This is of critical importance to help inform its future development. In this project, it will be necessary to design, develop and deploy "measurement software" across the world to collect meaningful data regarding Africa's network paths, server locations, websites, mobile technologies etc. Through this, the student will gain an expert understanding in how the Internet in Africa is deployed an operated. Using this expertise, the project will then progress to propose technologies and techniques to improve any issues found.

Return to top

Next-Generation Optical Access Network Management


Research group: Networks

The growth of video and high-speed packet-based business services, such as Ethernet private line, grid computing, and storage networking, have highlighted the need for customer-centric on-demand channel provisioning particularly in the access network. Recent advances in optical networking technology provide additional capacity. However, this new optical technology lacks the means of managing the optical resources efficiently in this highly dynamic environment. To address this problem, this research proposes to employ and extend Software Defined Networking to provide a flexible and scalable automated resource management framework for an optical access network. The approach involves exploiting novel algorithmic techniques to guarantee near-optimal performance, even under non-stationary and uncertain conditions.

Return to top

Intelligent Sensing for Patient Rehabilitation


Research group: Networks

The aim of this project is to devise a novel means of obtaining positional data for motor rehabilitation. In particular, multiple sensors in a garment are used to derive salient motion information based on how certain electrical characteristics change over time. However, we will do this in such a way that the exact location of each sensor on the patient’s body need not be known. Instead we have some information regarding the relative position of the sensors and anatomical constraints of body area of interest. This motion information is then processed and fed back to the patient to assist them in performing various therapeutic exercises.

Return to top

Improving the statistical basis of forensic evidence


Research group: Risk & Information Management

While DNA evidence has a (peculiarly undeserved) status as being ‘statistically sound’, other types of forensic evidence (Including fingerprints, footprints, shoeprints) do not enjoy such a status. Hence, whereas DNA experts are allowed to attach statistical assertions to the probative value of their evidence (such as ‘random match probabilities) the same is not true of experts in others areas of forensics. The hypothesis of this proposed research is that it is possible to improve the statistical basis for all types of forensic evidence (including DNA) by incorporating expert judgment using Bayesian methods. One of the objectives is to apply the work to provide a sound statistical basis for palynology evidence (this will be done in collaboration with a world-leading palynologist).

Return to top

Information Extraction and Computational Linguistics: A Case for Probabilistic Datalog


Research group: Risk & Information Management

Probabilistic Datalog (PDatalog) is a rule-based programming paradigm that provides a high-level data abstraction. PDatalog can be applied to information management tasks such as classification, summarisation, semantic (knowledge-based) retrieval, prediction and recommendation. This project aims at exploring the options to model methods and algorithms from information extraction and computational linguistics in PDatalog. The syntax and meaning of language can be captured in rules (onthologies), and the semantics of a text can be modelled as a set of facts and rules. The purpose of this project is to investigate the application of probabilistic reasoning to extract information and to reason about language. There are numerous challenges to be addressed. The main hypothesis is that many knowledge engineers (data analysts) can benefit from a high-level abstraction to model methods used for information extraction and in computational linguistics.

Return to top

The Robot Scientist: Active Machine Learning for Structured Models


Research group: Risk & Information Management

In traditional supervised machine learning, the human defines, collects, and annotates the data the machine should learn from. Thus the machine is a passive learner, with no ability to influence its training program. The result in many cases is dramatically sub-optimal learning, with much more data (and more importantly and expensively, more annotation) required than necessary. The alternative is active learning or optimal experimental design, where the machine performs more humanlike introspection, and decides the most informative data, explicitly asking the human teacher for the explanation (annotation) of that data. In this project, we will investigate cost-sensitive active learning models for structured domains such as multimedia. Technical challenges will be developing introspection models to estimate and trade-off the expected benefit of learning from different kinds of data and annotation, as well as developing efficient incremental learning algorithms to make this feasible. Possible application areas include computer vision/multimedia, medical diagnosis, risk management, etc.

Return to top

Long term football prediction using knowledge and data


Research group: Risk & Information Management

While Bayesian networks (which incorporate expert knowledge and data) have been used effectively to predict the outcomes of individual matches, there has been little effective work on longer term predictions such as predicting which league position a team will finish in at the start of a season. This project will investigate the use of Bayesian networks for this kind of longer term prediction

Return to top

Preserving empirical data in expert built Bayesian networks


Research group: Risk & Information Management

In the absence of empirical data we rely on expert judgment to build rich causal models. For such models the expert has to provide their own judgment about the probability of a variable conditioned on the parents (causes) of the variable. But in many cases, although we have no direct empirical data to inform these necessary conditional probability distributions we may have solid empirical data about the marginal probabilities of one or may variable or about P(A|B) where B is not a parent of A. While there are a small number of special solutions for this problem there is no generic strategy for ensuring that the empirical data is properly ‘preserved’ in the expert built model. This project will determine such a strategy.

Return to top

Research group: Risk & Information Management

Bayesian networks offer exciting potential to improve reasoning in complex legal arguments. The objective of this research is to determine the practicality of using Bayesian networks to model real-world cases involving forensic and other evidence

Return to top

The non-stop learner: Online Life-Long Machine Learning


Research group: Risk & Information Management

Traditional supervised machine learning attempts to solve any posed problem from scratch. Lifelong machine learning aims to provide the more humanlike ability to run for an extended period of time, addressing many diverse problems, and eventually learning general knowledge / skills that can be re-used to help improve performance at all tasks, and especially improve the efficiency of acquiring new skills. This project will focus on online algorithms for life-long learning. How can knowledge be restructured, re-used and synergistically combined when dealing with a realistic stream of tasks. Possible application areas include computer vision/multimedia, or robotics.

Return to top

Information Retrieval Models for Probabilistic Data Analysis


Research group: Risk & Information Management

Information Retrieval (IR) relies on probability theory. Retrieval models deliver a relevance-based ranking of retrieved objects, and many decades of research have shown that the independence assumption (often applied in probability theory) leads to sub-optimal retrieval quality. IR models incorporate DEPENDENCE ASSUMPTIONS, and therefore achieve good retrieval quality. The aim of this project is to transfer IR models to the general world of probability theory. The hypothesis is that modelling the dependence in probabilistic models (e.g. for health and law) can significantly improve the quality of models used for data analysis.

Return to top

Interaction methods for virtual reality and robotics involving sense of motion and touch


Research group: Robotics Engineering

Conventional virtual reality systems are based on generating realistic visual sensory modality, which is also the case for variety of telerobotics applications in which a user controls a robot or virtual avatar from a remote location. In this project novel methods of integration of sense of motion and touch in virtual reality and robotics applications will be investigated. You will develop novel robotics interfaces for generating simultaneous vestibular (sense of motion) and haptic (touch) illusions for improving immersiveness and enhance user engagement in VR and telerobotics applications. During the project you will learn how to build, programme and evaluate interactive robotic and computing systems and the results of this study will potentially induce the deployment of such systems into real life applications changing the ways we interact with machines. The project is led by Dr Ildar Farkhatdinov, e-mail: i.farkhatdinov@qmul.ac.uk, in co-operation with the robotics group led by Professor Kaspar Althoefer.

Return to top

Soft haptics - embedding interactive sense of touch in soft materials


Research group: Robotics Engineering

Haptic technologies can provide artificial computer controlled sense of touch to humans in robotics and virtual reality for various application sectors. In this project you will investigate and develop new technologies for providing sense of touch through integrating conventional haptic actuators with soft materials and pneumatic actuation. This will require design of novel mechatronic systems, their manufacturing and validation. Potential application and case studies will include wearable robotics and medical simulation. The project is led by Dr Ildar Farkhatdinov, e-mail: i.farkhatdinov@qmul.ac.uk, in co-operation with the robotics group led by Professor Kaspar Althoefer.

Return to top

Interaction methods for virtual reality and robotics involving sense of motion and touch


Research group: Robotics Engineering

Conventional virtual reality systems are based on generating realistic visual sensory modality, which is also the case for variety of telerobotics applications in which a user controls a robot or virtual avatar from a remote location. In this project novel methods of integration of sense of motion and touch in virtual reality and robotics applications will be investigated. You will develop novel robotics interfaces for generating simultaneous vestibular (sense of motion) and haptic (touch) illusions for improving immersiveness and enhance user engagement in VR and telerobotics applications. During the project you will learn how to build, programme and evaluate interactive robotic and computing systems and the results of this study will potentially induce the deployment of such systems into real life applications changing the ways we interact with machines. The project is led by Dr Ildar Farkhatdinov, e-mail: i.farkhatdinov@qmul.ac.uk, in co-operation with the robotics group led by Professor Kaspar Althoefer.

Return to top

Quantum Computation for Natural Language Processing


Research group: Theory

Vectors and their inner product geometry provide a mathematical basis for models of quantum mechanics. This setting has also been widely used to model natural language data. Herein, one builds vectors for words based on their co-occurrence frequencies in context. The geometric distances between these vectors represent similarity. Previous work --by van Rijsbergen, by Sordoni, and by Clark-Coecke-Sadrzadeh -- has shown how concepts from vector models of quantum mechanics can be carried over to vector models of language. These indeed provide a richer more general setting for natural language data, but it remains to show how advantagous are their faster computational methods. Might they result in speed ups in natural language processing, in the same way as they did for numerical factorization and sorting problems? This project aims to provide some answers.

Return to top

Higher-order Game Theory


Research group: Theory


This project aims to re-develop Game Theory using notions of higher-type computation, as in

All notions of Game Theory (such as player, game, strategy, equilibrium) can be recast in terms of higher-order constructions, or properties of such constructions. More details can be found in:

Some knowledge of a strongly typed functional language such as Haskell would be desirable.

Return to top

Programs and Proofs


Research group: Theory


This project aims to apply techniques of mathematical logic and proof theory in the extraction of correct programs from proofs. The main tools used are proof translations and proof transformations, e.g.

A good knowledge of propositional and predicate logic is necessary. For more papers in the areas see:

Return to top

Compositional Natural Language Methods for Information Retrieval


Research group: Theory

Information Retrieval and Natural Language Processing both rely on vector spaces and statistical measures as their underlying mathematical models. In Information Retrieval, queries and documents are represented by vectors, based on the number of times a term occurs in each, on the document and query length, and on collection-based statistics. These raw counts are turned into various different types of probabilities to smoothen the sparsities, scale down the effect of very frequent words, and promote the effect of important words. In Natural Language Processing, word vectors are obtained by counting co-occurrence in context; phrase and sentence vectors are obtained by composing the word vectors therein taking into account their grammatical structures. The vector space dimensions correspond to terms (contexts), and this feature space is different from the traditional IR case. Whereas Information Retrieval uses advanced engineering based statistical measures, Natural Language Processing is very good in using the linguistic structure and semantics of phrases and sentences. This project aims to bring these two fields together in such a way that they inherit the strong points of each other. As a result, from one point of view, Information Retrieval will become able to build vectors for queries and documents based on their internal linguistic structures. On the other hand, Natural Language Processing will become able to hire and apply advanced statistical measures. A rich application area is to retrieval in search engines and question answering.

Return to top

Quantum Computation for Natural Language Processing


Research group: Theory

Vectors and their inner product geometry provide a mathematical basis for models of quantum mechanics. This setting has also been widely used to model natural language data. Herein, one builds vectors for words based on their co-occurrence frequencies in context. The geometric distances between these vectors represent similarity. Previous work --by van Rijsbergen, by Sordoni, and by Clark-Coecke-Sadrzadeh -- has shown how concepts from vector models of quantum mechanics can be carried over to vector models of language. These indeed provide a richer more general setting for natural language data, but it remains to show how advantagous are their faster computational methods. Might they result in speed ups in natural language processing, in the same way as they did for numerical factorization and sorting problems? This project aims to provide some answers.

Return to top

The shape of natural language data


Research group: Theory

Distributional models of natural language build vectors for words based on their co-occurrences in context. Compositional distributional models extend these from words to phrases and sentences. These models have been very successful in detecting similarity and have found applications in various language tasks such as disambiguation and paraphrasing. But the general shape of the data remains a mystery. Do they, for example, adhere to linearity or might it be unavoidable to use non-linear concepts and algorithms? Existing studies indicate that in certain lower dimensional spaces linearity is the case. However, non-linear generalizations thereof, for instance those used by neural net formalisms, have provided certain better predictions. This project aims to explore the reality and provide a modular study. What is the real shape of natural language data? Would they differ for words as opposed to phrases and sentences? How much do we need to reduce dimensionality to arrive to non-linear models? Would any information be lost in this linear-to-non linear passage? If so, why non-linear methods perform better?

Return to top

Compositional Natural Language Methods for Information Retrieval


Research group: Theory

Information Retrieval and Natural Language Processing both rely on vector spaces and statistical measures as their underlying mathematical models. In Information Retrieval, queries and documents are represented by vectors, based on the number of times a term occurs in each, on the document and query length, and on collection-based statistics. These raw counts are turned into various different types of probabilities to smoothen the sparsities, scale down the effect of very frequent words, and promote the effect of important words. In Natural Language Processing, word vectors are obtained by counting co-occurrence in context; phrase and sentence vectors are obtained by composing the word vectors therein taking into account their grammatical structures. The vector space dimensions correspond to terms (contexts), and this feature space is different from the traditional IR case. Whereas Information Retrieval uses advanced engineering based statistical measures, Natural Language Processing is very good in using the linguistic structure and semantics of phrases and sentences. This project aims to bring these two fields together in such a way that they inherit the strong points of each other. As a result, from one point of view, Information Retrieval will become able to build vectors for queries and documents based on their internal linguistic structures. On the other hand, Natural Language Processing will become able to hire and apply advanced statistical measures. A rich application area is to retrieval in search engines and question answering.

Return to top

The shape of natural language data


Research group: Theory

Distributional models of natural language build vectors for words based on their co-occurrences in context. Compositional distributional models extend these from words to phrases and sentences. These models have been very successful in detecting similarity and have found applications in various language tasks such as disambiguation and paraphrasing. But the general shape of the data remains a mystery. Do they, for example, adhere to linearity or might it be unavoidable to use non-linear concepts and algorithms? Existing studies indicate that in certain lower dimensional spaces linearity is the case. However, non-linear generalizations thereof, for instance those used by neural net formalisms, have provided certain better predictions. This project aims to explore the reality and provide a modular study. What is the real shape of natural language data? Would they differ for words as opposed to phrases and sentences? How much do we need to reduce dimensionality to arrive to non-linear models? Would any information be lost in this linear-to-non linear passage? If so, why non-linear methods perform better?

Return to top

Software optimization and compilers


Research group: Theory

Production compilers such as LLVM, GCC, Intel C compiler, and Microsoft Visual Studio compiler can generate efficient code for common target architectures. However, it is challenging for compiler developer to keep up with the huge variety of new processor designs, spanning a spectrum of low-power and high-performance computing. The goal of this project is to design and implement a new compiler architecture that (a) guarantees correctness of generated code and (b) provides a way to explore the trade-off between compilation time and efficiency of generated code. If a programmer increases the time budget allotted for compilation, it will generate better code in terms of running time, code size, power consumption, energy efficiency, or other metrics. Ultimately, the generated code can take full advantage of the capabilities of new process designs, without the need to modify the compiler. The idea is to perform code generation and optimization using novel constraint solving techniques. This approach separates the concerns of software application logic, programming language semantics, architecture specification, and hardware performance modelling. Programmers, compiler writers, and hardware designers will benefit from this approach, which allows each of them to concentrate on aspects of the problem that match their expertise. A suggested starting point for this project is based on LLVM compiler and modern SMT Solver technology. Research can focus on theoretical or practical aspects of the project, depending on the interests and skills of the student.

Return to top

The non-stop learner: Online Life-Long Machine Learning


Research group: Vision

Traditional supervised machine learning attempts to solve any posed problem from scratch. Lifelong machine learning aims to provide the more humanlike ability to run for an extended period of time, addressing many diverse problems, and eventually learning general knowledge / skills that can be re-used to help improve performance at all tasks, and especially improve the efficiency of acquiring new skills. This project will focus on online algorithms for life-long learning. How can knowledge be restructured, re-used and synergistically combined when dealing with a realistic stream of tasks. Possible application areas include computer vision/multimedia, or robotics.

Return to top

Internet of Things for a Smart Planet


Research group: Vision

The Internet of Things consists of Smart Devices that are smart in the sense because they are active, digital, networked, can operate to some extent autonomously, are reconfigurable and have local control of the resources they need such as energy, data storage, etc. Currently the Internet of Things (IoT) predominantly focusses on internetworking things in urban areas to help enable the vision of a Smart City that is sustainable, energy efficient, etc. In contrast to Smart City IoTs that tend to support data exchange from Things in communication resource rich environments that are covered by multiple short-range and longer range wireless networks, in contrast exchanging data in non-urban Smart Planet IoTs is more challenging. This is because Things can be situated in harsh physical environments, often where the use of mainstream airborne wireless communication may not even be available and where the resources for Information and Communication Technology (ICT), and energy, are constrained. The primary objective is to research and validate the use of a system to support the Smart Planet data exchange requirements for openness, flexibility, scalability, resiliency and time-sensitivity.

Return to top

Privacy protection against spatial-temporal tracking


Research group: Vision

One of the main challenges for more widespread human use of IoT, which consists in part of more smart data sources and sinks that accompany humans when they move around, or are embedded in human environments, is data privacy for personal information acquired about individuals such as spatial-temporal tracks that perhaps detail our every movement. Increasingly, these tracks may not remain on a user’s mobile device where they could be protected via standard security mechanisms such as encryption but they need to be shared with location-based service providers and other third parties in order to access their services. Personal data privacy threats are that even if a pseudonym is used to identify a user’s actual identity, someone can be indirectly identified via their spatial-temporal data that is collected, i.e., via sensitive locations, e.g., where someone lives, and via someone’s spatial-temporal tracks being unique and repeatable. This project will investigate existing approaches to privacy protection against spatial-temporal tracking, and propose and validate new approaches.

Return to top

User-centered Security and Privacy for IoT use in eHealth Wearables and Wireless Sensor Networks


Research group: Vision

There is an increasing rich use of different types of IoT for eHealth, to monitor humans during daily-life that are self-driven and self-managed so that human subjects can be better informed to live and exercise healthier. This introduces new challenges for information security and privacy as such eHealth IoTs seriously invade users’ personal, social and public living spaces. Whilst, some of the security and privacy challenges for IoT eHealth monitoring have been identified, together with potential solutions, these have important limitations. One of the main limitations is that these lack a user-centered approach to secure how humans think and act, interact with other and interact with changing physical environments. This project will research and develop user-centered security and privacy methods and mechanisms for IoT in eHealth. These will enhance the behaviour of IoT eHealth systems to be better attuned to serve humans and to anticipate and respond to new and emerging security threats in constantly changing and complex physical and human environments.

Return to top

Socially aware cabinet robot


Research group: Vision

Social signal processing is an increasingly important element of design for affective computing applications. This project will look at developing and testing a software system that controls the configuration of a desktop storage unit depending on the expressions, engagement and gestures of the user, fused with other contextual data about the user and their environment. For example if the user is looking tired ( detected visually ) after a long day ( this is known as users diary of appointments is assessed) the system detects this user state and context and the drawer on the desktop opens to reveal the user MP3 players ready for a relaxing session. The project will develop students mathematical, programming, psychology and experimental skills.

Return to top

Deformable Stroke Models for Human Sketch Synthesis


Research group: Vision

Sketching comes naturally to humans. With the proliferation of touchscreens, we can now sketch effortlessly and ubiquitously by sweeping fingers on phones, tablets and smart watches. Studying free-hand sketches has thus become increasingly popular in recent years, with a wide spectrum of work addressing sketch recognition, sketch-based image retrieval, and sketching style and abstraction. While computers are approaching human level on recognizing free-hand sketches, their capability of synthesizing sketches, especially free-hand sketches, has not been fully explored. The main existing works on sketch synthesis are engineered specifically and exclusively for a single category: human faces. In this project, going beyond just one object category, we aim to introduce a generative data-driven model for free-hand sketch synthesis of diverse object categories. In contrast with prior art, (i) the model should be capable of capturing structural and appearance variations without the handcrafted structural prior, (ii) no purpose-built datasets should be required to learn from, but instead publicly available datasets of free-hand sketches are utilized, and (iii) the model should optimally fi ts free-hand strokes to an image via a detection process, thus capturing the specific structural and appearance variation of the image and performing synthesis in free-hand sketch style.

Return to top

Modelling Human Memory for Forensic Facial Sketch Matching


Research group: Vision

Facial sketch recognition is an important law enforcement tool for determining the identity of criminals where only an eyewitness account of the suspect is available. In this situation, a forensic sketch artist renders the face of the suspect by hand or with compositing software based on eyewitness description. The facial sketch is then disseminated in the media, but the crucial capability is to then identify the suspect by matching it against a photo mugshot database. This project investigates whether it is possible to model the human memory's forgetting process for faces, and whether such a model can be used to improve the performance of automated facial forensic sketch matching. Forensic facial sketch recognition is a key capability for law enforcement, but remains an unsolved problem. It is extremely challenging because there are three distinct contributors to the domain gap between forensic sketches and photos: The well studied sketch-photo modality gap, and the less studied gaps due to (i) the forgetting process of the eye-witness and (ii) their inability to elucidate their memory. The PhD student will have access to a database of 800 forensic sketches created at different time-delays, which is the largest such dataset to date.

Return to top

On-the-Fly Fine-Grained Sketch-based Image Retrieval


Research group: Vision

This project investigates the problem of fine-grained sketch-based image retrieval (SBIR), where free-hand human sketches are used as queries to perform instance-level retrieval of images. This is an extremely challenging task because (i) visual comparisons not only need to be fine-grained but also executed cross-domain, (ii) free-hand (finger) sketches are highly abstract, making fine-grained matching harder, and most importantly (iii) annotated cross-domain sketch-photo datasets required for training are scarce, challenging many state-of-the-art machine learning techniques. In this project, we will address all these challenges, providing a step towards the capabilities that would underpin a commercial sketch-based image retrieval application. In particular, we will investigate towards novel means of incorporating user feedback into the retrieval loop, with the aim of (i) eliminating variances of sketching skills among users, and (ii) improve overall retrieval accuracy.

Return to top

Magic driven visual hypothesis testing


Research group: Vision

Dynamic visual attention can be modelled computationally, but can it be used to predict the errors humans make in observing a sleight of hand magic trick? This project will allow a better understanding of the high level vs low level contributions in visual hypothesis formation through exploiting effects where the magician deliberately manipulates the cognitive context.

Return to top

The Robot Scientist: Active Machine Learning for Structured Models


Research group: Vision

In traditional supervised machine learning, the human defines, collects, and annotates the data the machine should learn from. Thus the machine is a passive learner, with no ability to influence its training program. The result in many cases is dramatically sub-optimal learning, with much more data (and more importantly and expensively, more annotation) required than necessary. The alternative is active learning or optimal experimental design, where the machine performs more humanlike introspection, and decides the most informative data, explicitly asking the human teacher for the explanation (annotation) of that data. In this project, we will investigate cost-sensitive active learning models for structured domains such as multimedia. Technical challenges will be developing introspection models to estimate and trade-off the expected benefit of learning from different kinds of data and annotation, as well as developing efficient incremental learning algorithms to make this feasible. Possible application areas include computer vision/multimedia, medical diagnosis, risk management, etc.

Return to top

The non-stop learner: Online Life-Long Machine Learning


Research group: Vision

Traditional supervised machine learning attempts to solve any posed problem from scratch. Lifelong machine learning aims to provide the more humanlike ability to run for an extended period of time, addressing many diverse problems, and eventually learning general knowledge / skills that can be re-used to help improve performance at all tasks, and especially improve the efficiency of acquiring new skills. This project will focus on online algorithms for life-long learning. How can knowledge be restructured, re-used and synergistically combined when dealing with a realistic stream of tasks. Possible application areas include computer vision/multimedia, or robotics.

Return to top

Subjective perception of scanned 3D models


Research group: Vision

The commercialization of depth-camera technologies (e.g. time-of-flight imaging) has made 3D data much more widely available. This includes close-range models of faces and objects, as well as large-scale scene models. There is great potential for this data to be used in cinema effects, virtual reality, game design, and HCI applications. However, it is not clear how best to process and render the raw 3D data, in order for it to be seamlessly merged with traditional video footage. This project will explore the perceptual aspects of this problem, using computational and psychophysical methods. In particular, 3D and head-mounted displays will be used, as well as eye-tracking and other sensor technologies. Strong programming skills, and a background in computer science, computer graphics, psychophysics, or computational neuroscience is required. There is scope for collaboration with the QMUL School of Experimental Psychology, and with visual effects companies, in this project.

Return to top

Physically-based modelling and rendering of 3D point-cloud data


Research group: Vision

3D data from depth-cameras and scanners is typically obtained in the form of semi-organized point clouds. This means that there are adjacency relationships between visual rays, from the viewpoint of each scan, and that the relative positions and orientations of the scans have been estimated. Large-scale point clouds, however, are often incomplete and inconsistent, owing to occlusions, overlaps, and viewpoint constraints. This project will investigate new geometric representations for point cloud data, with an emphasis on physically-based modelling and rendering. In particular, the problems of surface interpolation, re-sampling, physical consistency, and animation will be investigated. Strong programming skills, and a background in computer science, physics, or applied mathematics is required. There is scope for collaboration with the QMUL School of Geography, in this project.

Return to top

Deep Learning based Person Re-Identification


Research group: Vision

Person re-identification (Re-ID) is the problem of matching people across non-overlapping cameras views. Despite extensive efforts in the past decade, it remains an unsolved problem. This is because a person's appearance often changes dramatically across camera views due to changes in body pose, view angle, occlusion and illumination conditions. Recently, inspired by the success of deep neural networks, particularly deep Convoluational Neural Networks (CNNs) in various vision problems such as face verification, deep Re-ID models started to attract attention. However, unlike in other visual recognition problems, only modest success has been achieved. This is because Re-ID poses unique challenges to deep learning: apart from the difficulties in collecting large labelled datasets, there exist severe domain shift problems. Specifically, within each dataset, different camera views are drastically different causing cross-view domain shift; large cross-dataset domain shift also exists rendering the transfer learning between datasets difficult. In this project, we aim to develop novel deep neural network architectures and loss functions tailor made for the Re-ID problem, as well as novel deep transfer learning models to overcome the problem of lacking sufficient training data. Dr Tao Xiang has been a world-leading expert on person Re-ID (see http://www.eecs.qmul.ac.uk/~txiang/publications.html), and the research group has one of the most powerful GPU server clusters in the UK academia, thus providing a perfect environment for this project to succeed.

Return to top

Learning from weakly labelled data for automatically generating image captions


Research group: Vision

One of the key human abilities that researchers have striven to emulate is to understand the content of images and describe it using language. Solving this problem not only advances fields such as computer vision, multi-media and machine learning, it also has many applications such as semantic image search, and providing image interpretation for the visually impaired. Existing models are learned from strongly aligned image-text pairs. However such a fully supervised approach based on strongly labelled data suffers from the problem of lacking sufficient annotated data, given the complexity and diversity of unconstrained image and language content. On the other hand, on the internet there exist unlimited amount of unlabelled data in both modalities, as well as weakly labelled image-text pairs where images and text are loosely aligned. An unrealised vision is thus to have a unified framework which can seamlessly exploit at the Internet scale the mined image and text data of all different forms, ranging from fully labelled to completely unlabelled. To this end, this PhD project aims to develop a novel framework that learns a translation model between images and a given language by leveraging weakly labelled data as well as fully labelled data whenever they are available. This framework is one step closer towards the grand challenge of life-long learning by which a computer can automatically and continuously learn by mining information from big data on the web, growing its understanding in an open-ended way, and eventually reproducing the visual perception and translation perception ability of humans.

Return to top

The Robot Scientist: Active Machine Learning for Structured Models


Research group: Vision

In traditional supervised machine learning, the human defines, collects, and annotates the data the machine should learn from. Thus the machine is a passive learner, with no ability to influence its training program. The result in many cases is dramatically sub-optimal learning, with much more data (and more importantly and expensively, more annotation) required than necessary. The alternative is active learning or optimal experimental design, where the machine performs more humanlike introspection, and decides the most informative data, explicitly asking the human teacher for the explanation (annotation) of that data. In this project, we will investigate cost-sensitive active learning models for structured domains such as multimedia. Technical challenges will be developing introspection models to estimate and trade-off the expected benefit of learning from different kinds of data and annotation, as well as developing efficient incremental learning algorithms to make this feasible. Possible application areas include computer vision/multimedia, medical diagnosis, risk management, etc.

Return to top