Research

With the realisation of new generations of network technologies, the integration and interoperation of multiple heterogeneous networks (such as cellular, WiFi and vehicle-to-vehicle networks, Internet-of-things) supporting mobile communications and AI services and applications will be of great importance. Developing AI and machine learning (ML) techniques to design, control and optimise future networks are also important. The Group is concerned with research leading to the development of techniques to enhance the responsiveness of new technologies, control protocols and mechanisms and systems as a whole when facing dynamic traffic changes, user mobility, a variety of user requirements and the deployment of new services. In addition to using various mathematical techniques such as optimization, and stochastic and statistical analysis, advanced AI and ML approaches are also investigated and developed to solve complex network problems.

Academics: Kin K. Leung, Bruno Clerckx, Stefan Vlaski

Array Communications and Array Processing have evolved into a well-established research area moving from old diversity systems, conventional direction nulling and phased-arrays to space-time communications, advanced superresolution direction finding algorithms and superresolution beamformers. The idea of combining Arrays with Communication Systems has the potential of providing more powerful wireless communication systems where both space and time information is exploited. 

Academics: Athanassios Manikas

Image and Signal Processing has a lot to offer to Biomedical sciences: Biomedical data analysis may be used to make explicit information that is implicit in the data, thus help us understand better how the human body works and also help the clinicians with diagnosis and visualisation of their data. The Group has active projects concerning the analysis of EEG and MEG data, 3D MRI data related to cancer research, microarray data concerning gene expression, and work combining computer graphics with image processing technologies for plastic surgery planning.

We are also pioneers of so called Hearables – in ear sensing of EEG and vital signs (ear-EEG, ear-ECG) – with applications to continuous 24/7 unobtrusive sensing in real word, especially related to sleep quality, fatigue and drowsiness. Current projects include the use of our ultra-wearable Hearables in dementia research and in head trauma and concussions (horse riding). We hold patents in this area.

The CSP group hosts the Smart Environment Lab (SEL), a Faculty facility that offers excellent state-of-the-art facilities for data capture and processing from a variety of sensors.

Academics: Danilo Mandic, Tania Stathaki

Communication and information theory deals with fundamental limits of the processing, transmission, storage, and use of information. It encompasses theoretical and practical aspects of coding, communications, cryptography, data processing and learning.

We study the fundamental limits of communication systems and derive coding and communication strategies that achieve those fundamental limits.

One area of research is on identifying the fundamental limits of MIMO wireless networks. We aim at understanding how to manage interference in the most efficient way and how to make the best use of multiple antennas and channel state information. This finds direct applications in all terrestrial (e.g., 5G and beyond) and non-terrestrial communication (e.g. satellite) systems.

Another area of research is coding theory and their applications in communication and post-quantum cryptography. Post-quantum cryptography, typically based on the hardness of coding and lattice problems, addresses secure communications in a post-quantum era when conventional public-key cryptosytems would be broken by quantum computing attacks.

Academics: Bruno Clerckx, Cong Ling, Geoffrey Li

Computational imaging refers to a set of imaging techniques that combine the design of the hardware layer (e.g., optical components, illumination, sensors, and devices) with signal processing techniques in order to go beyond physical limitations of traditional optical systems and achieve novel imaging capabilities that one could not with traditional imaging methods. In computational imaging, computation plays an integral role in the image formation process, for this reason, this research area is intimately related to sampling theory and aspects of sparse sampling.

The group collaborates with many institutions and other departments on this topic. In particular, with the bio-engineering department, the group is active in using computational imaging methods to produce functional images of the brain at cellular resolution using two-photon microscopes. The group interacts also with the National Gallery to develop new computational imaging techniques to analyse Old-Master paintings. It is also involved in 3-D imaging using single pixel or single-photon time-of-flight detectors.

Academics: Pier Luigi Dragotti, Ayush Bhandari, Wei Dai

The use of computational methods in subjects archaeological and in matters related to paintings central to the collaborative effort with the University of Athens, the archaeological excavations in the prehistoric site of Akrotiri, Thera,  and University College, London (UCL), Institute of Archaeology. The work has been in place since 1985 with the restoration of Florentine Renaissance wall paintings. Current efforts with considerable success are directed to a broad range of objectives amongst which included are the computational reconstruction of buildings,  the assembly of wall paintings from small and degraded fragments, the identification of symbols in degraded papyri, the identification of scribes in the UCL Lahun collection of papyri and others.

Academics: Tania Stathaki, Tony Constantinides

Computer vision is closely associated with image processing and patter recognition. In the simplest terms, computer vision aspires to make computers reason on the content of digital images. The Group has active research projects on cognitive vision that combines machine learning and computer vision, networks of cameras that cooperate to track moving objects, 3D reconstruction of objects for face modelling and recognition, and on more fundamental aspects of how the human visual system works.

Academics: Tania Stathaki

The proliferation and democratization of technology has resulted in the dispersion of data and computational resources. Access is controlled by independent users or entities, which are generically referred to as “agents”. Examples are sensors measuring weather data, mobile devices holding user data, or health care providers holding patient data. Agents are reluctant to share raw data due to concerns around privacy, communication, and power constraints. This prevents central aggregation of data and use of classical machine learning or signal processing frameworks. This dichotomy motivates the development of algorithms which allow agents to learn cooperatively without exchanging raw data. Motivated by these considerations, we develop new algorithms for distributed learning, optimization and signal processing, which are able to match the performance of centralized benchmarks while avoiding the aggregation of raw data. These structures are more resilient, more communication efficient and more private. 

Academics: Stefan Vlaski 

The Financial Signal Processing (FSP) Lab was created in 2014 with a vision of bringing professionals from academia and industry together to promote research in quantitative finance using engineering tools, with a special focus on signal processing and optimisation techniques.  The Lab is led by A. G. Constantinides, and co-directed by D. P. Mandic and the current team of the FSP lab includes professionals from both academia and industry including Professors and Senior Lecturers along with Managing Directors  and CEOs. Current applications include the modelling of financial decisions as recommender systems, Deep Reinforcement Learning for Finance, heavy tailed probabilistic mixture models (Elliptical Mixture Models), and Big Data Approaches to financial modelling.

Academics: Tony Constantinides, Danilo Mandic

Intelligent Communications 

Traditionally, communication systems are designed based on certain mathematical models and are hard to address issues such as non-linear distortion, robustness to unknown/unpredictable impairments, and semantic extraction and transmission. As pioneering researchers, the faculty in this group introduced DL in wireless communications to deal with these issues.  Deep neural networks were first used to jointly estimate the channels and detect the modulated signals successfully, which revealed that DL can mitigate both interference and non-linear distortion and becomes a popular research topic in wireless communications nowadays.  With the help of deep neural networks, text-semantic communication system was developed for the first time, which significantly outperforms the traditional ones. The group also investigated theoretical and practical issues of deep reinforcement learning for wireless resource allocation and wireless communications for federated learning.  Broadly speaking, the group develops intelligent wireless systems where tailored machine learning techniques are developed to either optimize those systems, including communications, sensing, power transfer or integrated version of them, typically for large scale scenarios or subject to highly nonlinear behaviors, or to examine how existing blocks of communication systems can be replaced by neural network architectures. 

Academics: Geoffrey Li, Bruno Clerckx 

The group are pioneers in both theoretical and applied aspects Recurrent Neural Networks (RNNs), including the first research monograph in this area (Wiley 2001). Ongoing work includes algorithms for Tensors for Big Data applications, Deep Neural Neural Networks, and Reservoir Computing, with seminal work on Tensor Networks for Dimensionality Reduction and Large Data Optimisation published as a two-volume monograph by Now Publishers in 2017 and 2018. Current work focuses on using super-compression associated with tensor-based approaches to reduce the dimensionality of DNNs, and to equip them with enhanced interpretability and explainability.  We also consider general mixture models, including Elliptical Mixture Models and recommender systems. Applications include Big Data for Finance, RNNs and Deep Learning for wearable sensors and imaging, and Machine Intelligence for Smart Grid.

Academics: Danilo Mandic, Tania Stathaki, Tony Constantinides

Image Processing encompasses a variety of techniques applied to digital images in the broadest sense of the word: optical images, hyperspectral images captured by satellites orbiting the Earth, 3D seismic images of the crust of the Earth, 3D tomographic images of the human body, as well as video sequences. The Group has active research on image fusion, enhancement, restoration, texture and shape analysis, object recognition, invariant feature construction, colour analysis etc.

Recently it has been more and more involved in the use of deep neural networks to solve inverse imaging problems as well as for image resolution enhancement and fusion with applications that span many domains including medical imaging.

Academics: Pier Luigi Dragotti, Tania Stathaki

Many problems in machine learning, signal processing and communications are fundamentally optimisation problems. We look for neural network parameterisations that fit the data best, signal reconstructions that minimise the mean-squared error, or communication systems that yield the smallest possible bit-error rate. As these systems grow in complexity, so do the underlying optimisation problems, necessitating the development of more sophisticated optimisation algorithms and performance analysis. Motivated by these developments, we have in recent years made contributions to distributed and stochastic optimisation, and provable non-convex optimisation. 

Academics: Stefan Vlaski 

With the recent climatic changes, the environment is at the forefront of public concern. Earth observation data coming from satellites orbiting the Earth may be combined with ground collected data, map information and other sources to help us monitor the state of the environment, create hazard maps for possible natural disasters, forecast and monitor events like landslides and floods, as well as manage resources and recommend actions. The group has a lot of experience in such research projects. The Group is also involved in analysing data from the Insight mission on Mars. 

Academics: Tania Stathaki, Danilo Mandic

The notion of sparsity, namely the idea that the essential information contained in a signal can be represented with a small number of significant components, is widespread in signal processing and data analysis in general. Great progress for example in image compression and enhancement has been obtained by modeling signals as sparse in an appropriate domain including the wavelet and frequency domains. The understanding that sparsity can be used to drive directly the information acquisition process is instead much more recent.

The group has years of experience in sparse signal representation, sampling based on sparsity models and applications in sparse inference, compression, super-resolution and tracking. Current research projects are in the area of dictionary learning for sparse representation, construction of sampling matrices/operators, finite rate of innovation sampling and a wide range of applications from estimation of diffusion fields, to imaging and neuroscience as well as channel estimation and sensing.

Working with Defence Science and Technology Laboratory (Dstl) and leading companies in defence industry, the group is also active in using super-resolution, and sparse signal processing methods to electromagnetic environment situation awareness.

Academics: Ayush Bhandari, Wei Dai, Pier Luigi Dragotti

Our goal is to study and improve human and machine hearing. The research of the Speech and Audio Processing Lab currently addresses single- and multi-channel acoustic systems, and speech processing for hearing aids and speech recognition systems. The Lab is a member of Imperial’s 'Natural and Machine Hearing’ initiative. The key technical topics in acoustic signal processing include microphone array beamforming, direction-of-arrival estimation, acoustic source tracking, acoustic imaging and acoustic SLAM. Our work on speech processing includes the study of aphasia following stroke and associated work on lesion mapping from the voice, binaural speech enhancement, speech dereverberation and audio-visual data fusion. Key recent partnerships are with University College London for work on hearing devices and we work fruitfully with several industry partners.

Academics: Mike Brookes, Patrick Naylor

The team research effort is directed towards the development of design techniques for fixed and adaptive parameter digital filters for applications including adaptive prediction, noise cancellation, system identification and equalisation. Moreover, it looks into implementation issues on a range of platforms (FPGA, DSP Chips, ASICs) and their application in a wide range of statistical signal processing problems. Particular emphasis is on multidimensional adaptive filters and ways of dealing with complex and quaternion noncircular (improper) signals. Applications include Signal Processing for Smart Grid and self-interference mitigation in full duplex transceivers in 5G communications. The group has granted patents in both these areas.

Academics: Danilo Mandic, Tony Constantinides

The highly successful introduction and rapid growth of mobile internet and wireless networks has re-emphasized the need for the efficient use of the limited bandwidth that is available. The activity of the group in wireless communications and signal processing is, in one way or another, concerned with the research into techniques to improve the spectral efficiency and energy efficiency, to cope with the massive increase of the number of devices, to boost the reliability of multi-user communication in the presence of noise, interference and over fading channels, and to invent the future electromagnetic environment. This group pioneered orthogonal frequency division multiplexing (OFDM) for wireless communications, which has been used in almost all wireless communication systems nowadays. We develop innovative communications strategies, and novel signal processing, optimization and machine learning tools and applications for wireless systems, networks and standards (5G, IoT, satellite, etc).

Specific topics include MIMO and multi-antenna signal processing, rate-splitting, robust interference management, modulation and coding, multiple access, cache-aided wireless communications, multi-user and massive MIMO, millimeter-wave and higher frequency bands. We developed the first MIMO-OFDM system in the world in 1990’s.

We are also heavily involved in understanding how wireless can be used not only for communications but also for other applications such as wireless power transfer, wireless information and power transmission, sensing, radar, localization, and how to make the best use of the spectrum, radiowaves and infrastructure to enable all those applications. Recent activities also include Integrated Sensing and Communications (ISAC) and Reconfigurable Intelligent Surfaces (RIS) to shape the Electromagnetic Environment and engineer the wireless channel, and find new potential in future very large scale antenna systems, such as ultra-massive MIMO.  To design those wireless systems, we can extensive use of communication / information theory, signal processing, machine learning, convex and non-convex optimisation, and microwave

We are involved in several research projects funded by UK research councils, Defence Science and Technology Laboratory (Dstl), EU and U.S. programmes and maintain close links with industries. Emphasis is put on theoretical algorithm developments but also on prototyping and experimentation, for both civilian and defence applications.

Academics: Bruno Clerckx, Kin K. Leung, Cong Ling, Geoffrey Li