Nature Inspired Computing and Engineering Research Group

We explore the use of modern machine learning algorithms such as deep neural networks for a wide spectrum of visual problems. The aim is to allow machines to see and understand the world around them. A computer vision system should be able to detect, track, identify and analyse the visual objects in an image or a video.

Computer vision technology has many practical applications that benefit society. For example, one can use computer vision models to help in looking after the wellbeing of elderly people who live alone in a house. We can also use computer vision algorithms to identify early-stage cancers by detecting subtle changes in tissues or cells. Autonomous driving systems also use many visual sensors to detect and analyse pedestrians, traffic signs and vehicles on the road, which provides rich information for decision-making.

We aim to develop cutting-edge computer vision algorithms in conjunction with modern machine learning approaches for solving a wide range of real problems in our life.

Contact us at nicequery@surrey.ac.uk if you'd like to find out more about our research in computer vision.

We research privacy-preserving AI techniques such as adversarial techniques for local anonymisation algorithms that reduce the risk of re-identification through linkage with other identifiable datasets. Federated learning seeks to make data ownership and provenance first-class concepts of learning and analytics systems through the principle of data minimization, as applied to aggregations.

Research into multi-objective evolutionary federated learning aims to minimise the model update communication cost; we are pushing the boundaries in this emerging field, with an emphasis on how privacy technologies may be combined in real-world systems. Our members also research interpretability of deep learning networks through feature interpretability for experiment design.

Contact us at nicequery@surrey.ac.uk if you'd like to find out more about our research in explainable and trustworthy AI.

With human-like and neural computing we take inspiration from how humans solve information processing problems and try to make machines perform in a similar way. This inspiration could come from human behaviour studies in psychology, or from neuroscience studies of how neurons process information at a lower level. For example, a human-like approach to video understanding can use interpretable models, where the computer can report why it recognised e.g. “handbag theft” in a given video. The computer would be able to identify the thief’s hand grabbing a bag, and who was the original owner of the bag, and would be able to explain why it is “theft” and not “giving” by pointing to the gaze direction of the actors.

In neuroscience-inspired studies, the behaviour of biological neural networks is simulated using computer models. Different methods exist to conduct such simulations, which also vary in their biological realism. For instance, spike-based neural networks (SNNs) have many benefits including energy-efficient and low-latency computation. It is currently an active research question how to create and develop functional SNNs, but the biological brain serves as an excellent inspiration. Along those lines, network cascades found in brain function (neuronal avalanches) provide a design paradigm for logic and learning under statistical learning models, such as the linear threshold model (LTM), which allow the study of spatial constraints and criticality to efficiency.

We are also a founding member of the EPSRC Network+ on Human-Like Computing (HLC) which is currently supporting the development of a UK-wide multi-disciplinary community of AI and cognitive science researchers.

Contact us at nicequery@surrey.ac.uk if you'd like to find out more about our research in human-like and neural computing.

Machine learning is a fundamental branch of artificial intelligence (AI) which investigates methods for modelling and performing predictions. It is also the building block for nearly all the research activities of our group. We regularly design and implement machine learning methods and investigate effective ways to train and evaluate these, with techniques including neural, evolutionary, reinforcement, statistical and logic-based machine learning and reasoning methods, and the integration of these approaches.

Integrating learning and reasoning constitutes one of the key open questions in AI and holds the potential of addressing many of the shortcomings of contemporary AI approaches, including the black-box nature and the brittleness of deep learning. Some of our more specialist methods include Monte Carlo sampling methods and the optimal transport theory, as well as novel logic-based and relational machine learning algorithms such as Meta-Interpretive Learning, and statistical relational learning algorithms such as MetaBayes.

The real-world problems we have applied these techniques to are in areas including healthcare, natural language processing, optimisation in transport networks, and metadata extraction from longitudinal social science questionnaires.

We also employed modern machine learning methods for various biomedical applications, such as for the diagnosis of Alzheimer’s disease and more generally neurodegenerative diseases.

Contact us at nicequery@surrey.ac.uk if you'd like to find out more about our research in machine learning and reasoning.

Optimisation is a subject transversal to artificial intelligence and more generally to engineering that lies at the intersection of mathematics and computer science and it appears in a number of topics across the activities of our group. For example, optimisation algorithms are needed to train machine learning components, to design robot controllers, to predict the outcome of an intervention in a complex system, to optimise a series of decisions in healthcare or security, and to address many computer vision and data science problems.

We use, develop, and design optimisation algorithms in various contexts. Our expertise ranges from exact algorithms, such as gradient-based methods, to nature inspired metaheuristics, such as evolutionary and swarm intelligence algorithms.

Contact us at nicequery@surrey.ac.uk if you'd like to find out more about our research in optimisation algorithms.

Biological systems are usually highly complex systems where cells interact with other cells, as well as with their spatial environment. Technological advances, such as in gene sequencing or brain imaging, allow researchers to collect enormous amounts of highly detailed information across different spatial and temporal scales. Computational approaches are very well-suited to visualise, simplify and model such data; this allows us to gain insights into the underlying rules of phenomena such as cancer, neural development, ecological networks, and many others.

We develop AI methods (declarative, relational and rule-based machine learning, deep reinforcement learning) for prediction and control in biological networks, with application to targeted therapeutics. For example, we have recently developed AI methods (using deep reinforcement learning) that demonstrate successful control of gene regulatory networks, where the target states comprise a tiny fraction of the state space (metastatic melanoma network with 2^70 states).

Our relational learning and text-mining approach for automated discovery of food-webs is regarded as the first successful application of machine learning from large scale agricultural data and the learned trophic links were confirmed by subsequent empirical studies (using DNA analysis) and were published in high impact scientific journals.

We developed the high-performance software BioDynaMo to model complex biological systems, including biologically realistic neural development. This approach has been used to model brain development as well as multiple biomedically relevant use cases, such as cancer growth and fibrosis.

Contact us at nicequery@surrey.ac.uk if you'd like to find out more about our research in systems biology and bioinformatics.