Centre for Computational Neuroscience and Robotics - CCNR

be.AI - Project Ideas

Supervisors have formulated project proposals that they have identified. They are listed below and can be used as the basis of your own PhD project proposal. Please click on the titles to see more details.

Understanding human time perception and memory for applications in human health and AI interaction

Supervisor: Warrick Roseboom, with potential co-supervisors Prof. Chris Bird (Sussex), Dr Sam Berens (Sussex), Dr Kaoru Amano (University of Tokyo, Japan), Dr Zafeirios Fountas (Huawei/UCL)

In the Time Perception group at Sussex, we are working to understand the cognitive and neural foundations of human time perception and memory for applications in health (e.g. digital memory augmentation in dementia) and human interactions with artificial systems (see e.g. TIMESTORM project; http://timestorm.eu). Potential projects could come from a cognitive neuroscience direction, focused on building foundation knowledge of the cognitive and neural bases of human time perception and memory, or instead focus on building AI systems that are enhanced by human-like temporal abilities – and any mix in between. Required skills depend on topic and we welcome proposals across disciplines from cognitive science to applications in machine learning. See below for examples of recent work that we seek to extend on:

Apply for a PhD in Informatics

[1] Roseboom, W., Fountas, Z., Nikiforou, K., Bhowmik, D., Shanahan, M. & Seth, A.K. (2019). Activity in perceptual classification networks as a basis for human subjective time perception. Nature Communications, 10, 267. doi: 10.1038/s41467-018-08194-7

[2] Sherman, M.T., Fountas, Z., Seth, A.K. & Roseboom, W. Accumulation of salient events in sensory cortex predicts subjective time. bioRxiv. doi: 10.1101/2020.01.09.900423

[3] Fountas, Z., Sylaidi, A., Nikiforou, K., Seth, A.K., Shanahan, M., & Roseboom, W. A predictive processing model of episodic memory and time perception. bioRxiv. doi: 10.1101/2020.02.17.953133

[4] Suárez-Pinilla, M., Nikiforou, K., Fountas, Z., Seth, A.K. & Roseboom, W. (2019). Perceptual content, not physiological signals, determines perceived duration when viewing dynamic, natural scenes. Collabra: Psychology, 5(1), 55. doi: 10.1525/collabra.234

[5] Zakharov, A., Crosby, M., Fountas, Z. (2020). Episodic Memory for Learning Subjective-Timescale Models. arXiv. arXiv:2010.01430v1

Risky decision-making: the neural computations underlying threat conflict

Supervisors: Kevin Staras, Michael Chrossley, and one of Andy Philippides, James Knight, or Thomas Nowotny

How do animals select appropriate actions when faced with conflicting threats, for example, predation and starvation? With increased hunger, a reprioritization of behaviours towards food-finding must be balanced against an elevated predation risk. The behavioural and neural mechanisms that underlie this are fascinating but completely uncharacterized. This project will combine the important simple-system model Lymnaea (Nat Commun 2016, 7:11793; Sci Adv 2018, 4:eaau9180) with state-of-the-art behavioural and neural activity assays, as well as computational modelling. This project will suit a student interested in developing skills in cellular neuroscience, cutting-edge behavioural assays, and circuit modelling approaches, to determine how animals balance risk and survival.

Apply for a PhD in Neuroscience

Understanding robot-animal interaction for autonomous navigation

Supervisors: Bao Kha Nguyen and Yanan Li

Autonomous navigation in dynamic environments has been extensively studied. Most of existing works model the robot’s environments based on the surrounding objects’ spatial locations or the temporal evolution of their movements. However, if a mobile robot navigates on an open farm surrounded by animals, these animals will likely change their movements according to the robot’s. In this sense, the bilateral adaptation between the robot and the animals needs to be investigated. This project will analyse animal behaviours and exploit learning algorithms to predict the animals’ movements in response to the robot’s. It will in turn enable the robot to effectively respond to and even affect the animals’ movements. This research will be useful for not only autonomous navigation of robots in dynamic environments but also for autonomous herding and management of animals.

Apply for a PhD in Engineering

Impact of behaviour and feedback projections on early visual processing

Supervisors: Dr Sylvia Schröder, Prof Leon Lagnado

Visual processing is not simply a sequential process extracting first simple, then complex visual attributes. Instead even early stages of visual processing integrate information about the internal and behavioural state of the animal and information from higher visual centres. We want to investigate the purpose and the mechanisms of this multi-faceted integration in the early visual system. How do locomotion, arousal and other behavioural contexts (e.g. visual decision making) affect visual responses in the retina and superior colliculus of the mouse? What is the advantage of integrating behavioural information for visual processing and visually guided behaviours?

Experimental techniques: two-photon imaging, electrophysiology using silicon probes (Neuropixels), opto-/chemogenetics

Apply for a PhD in Neuroscience

[1] Schröder, S., Steinmetz, N.A., Krumin, M., Pachitariu, M., Rizzi, M., Lagnado, L., Harris, K.D., and Carandini, M. (2020). Arousal Modulates Retinal Output. Neuron 107, 487-495.e9.

[2] Liang, L., Fratzl, A., Reggiani, J.D.S., El Mansour, O., Chen, C., and Andermann, M.L. (2020). Retinal Inputs to the Thalamus Are Selectively Gated by Arousal. Curr. Biol.

[3] McGinley, M.J., Vinck, M., Reimer, J., Batista-Brito, R., Zagha, E., Cadwell, C.R., Tolias, A.S., Cardin, J.A., and McCormick, D.A. (2015). Waking State: Rapid Variations Modulate Neural and Behavioral Responses. Neuron 87, 1143–1161.

Plasticity of visual computations

Supervisors: Prof. Leon Lagnado and Prof. Jeremy Niven

The synaptic connections that transfer information between neurons are likely to be crucial to the plasticity of computations in the brain but we have surprisingly little understanding of how. We study this problem in zebrafish, where turning decisions driven by vision are modulated by information arriving through other senses, such as smell. We use a combination of experiment (multiphoton imaging and behaviour), theory (information and signal detection) and modelling (biophysical) to understand how the plasticity of synapses in adjusts the computations that visual circuits carry out and the behaviours they drive. A number of projects can be offered depending on the applicants background e.g How does neuromodulation alter the efficiency with which synapses transfer information and how does this alter visually-driven motor decisions?

Apply for a PhD in Neuroscience

[1] Jose Moya-Diaz, Ben James, Federico Esposti, Jamie Johnston, Leon Lagnado (2021). Diurnal modulation of multivesicular release controls the efficiency of information transmission at a sensory synapse. bioRxiv doi: 10.1101/2021.09.12.459944.

[2] James, B., Darnet L., Moya-Diaz, J., Seibel, S-H., Lagnado, L. (2019). An amplitude code transmits information at a visual synapse. Nature Neuroscience, 22(7):1140-1147

Andreas M. Kist, Ruben Portugues, (2019). Optomotor Swimming in Larval Zebrafish Is Driven by Global Whole-Field Visual Motion and Local Light-Dark Transitions. Cell Reports, 29(3),659-670.

Understanding crowd psychology and safety through wearable technologies

Supervisors: Prof John Drury, Prof Daniel Roggen

Analysing and modelling the behaviour of people in crowds can help safely manage live events, improve response to emergencies, and inform the design of public spaces [1]. Wearable and mobile sensing now allows us to sense behaviour "in the wild" and in real-time [2] [3] [4] [5]. This project would bring together novel wearable and mobile sensing with concepts from group psychology to understand co-action and coordinated behaviour in pedestrian flow. The ideal student will have strong computing skills and a background in computer science, maths, physics or psychology. They will be interested in research at the interface between psychology and computing.

Apply for a PhD in Psychology

[1] Kleinmeier, Köster and Drury, Agent-based simulation of collective cooperation: from experiment to model. J. R. Soc. Interface, 2020

[2] Wirz et al. Probing crowd density through smartphones in city-scale mass gatherings. EPJ Data Science, 2(5):1-24, 2013.

[3] Wang et al. Enabling reproducible research in sensor-based transportation mode recognition with the Sussex-Huawei dataset. IEEE Access, 2019

[4] Chavarriaga et al. The Opportunity challenge: A benchmark database for on-body sensor-based activity recognition. Pattern Recognition Letters, 2013

[5] Templeton, A., Drury, J., Philippides, A. (2018). Walking together: Behavioural signatures of psychological crowds. Royal Society Open Science 5, 180172.

Probabilistic inference and control in animals and machines

Supervisors: Dr Christopher Buckley, and one or more of Prof Anil Seth, Prof Leon Lagnado, Dr Arash Moradinegade Dizqah

Animals are able to thrive in noisy and uncertain environments. Converging theory in the brain sciences suggests that they achieve this by operating as probabilistic inference machines [1,2,3]. On this view perception, action and learning can all be understood as a minimisation of the divergence between an inferred distribution over environmental states and a desired target distribution. This information theoretic starting point underpins modern algorithms in machine learning (e.g., intrinsic measures, maximum entropy reinforcement learning), current theory in the cognitive sciences,(e.g., active inference), and process theories in neuroscience (e.g., predictive coding and optimal control). While the major focus of the project will be on modelling and theory in this area, we particularly encourage proposals that demonstrate an ambition to ground ideas in either experimental neuroscience, machine learning practice or robotics. Applicants should have a background in a quantitative science. Experience with math and programming is essential. This project will also involve potential research exchanges with the Center for Human Nature, Artificial Intelligence, and Neuroscience (CHAINO) Hokkaido University, Japan.

Apply for a PhD in Informatics

[1] Pouget, A., Beck, J., Ma, W. et al. Probabilistic brains: knowns and unknowns. Nat Neurosci 16, 1170–1178 (2013).

[2] Karl J. Friston, Marco Lin, Christopher D. Frith, Giovanni Pezzulo, J. Allan Hobson, and Sasha Ondobaka. Active Inference, Curiosity and Insight. Neural Computation 2017 29:10, 2633-2683

[3] A. Tschantz, M. Baltieri, A. K. Seth and C. L. Buckley, "Scaling Active Inference," 2020 International Joint Conference on Neural Networks (IJCNN), Glasgow, United Kingdom, 2020, pp. 1-8

Using AI to improve autonomy in laser-based surgical robots

Supervisors: Rodrigo Aviles and Elizabeth Rendon

Future autonomous surgical robots will have the ability to “see”, “think” and “act” without active human intervention to achieve a predetermined surgical goal safely and effectively. The “robot’s surgical skill” consists of its ability to first map its perception (i.e. sensory inputs) to an estimated environmental state, and then map that estimate to a future action ( i.e robotic output) in the most efficient way possible. Machine learning (ML) is proposed as means to control the actions of autonomous devices [1]. Appropriately trained algorithms can enable robots presented with novel yet similar data, to predict an outcome, and then achieve its task in real time based upon its “experience”. If the sensory stream is of comparable fidelity to human senses, such analytical algorithms will demonstrate superiority over human perception.

This project will investigate types of machine learning algorithms relevant to train autonomous surgical devices. Combinations of supervised, unsupervised and reinforcement ML techniques will be investigated together with the development of advanced computer vision analysis and recognition models, to produce an autonomous device with the versatility required to perform a range of soft-tissue surgical procedures [2] using laser-based ablation for tumour resection.

Apply for a PhD in Engineering

[1] Kassahun, Y., Yu, B., Tibebu, A.T., Stoyanov, D., Giannarou, S., Metzen, J.H., Vander Poorten, E., 2016. Surgical robotics beyond enhanced dexterity instrumentation: a survey of machine learning techniques and their role in intelligent and autonomous surgical actions. International Journal of Computer Assisted Radiology and Surgery 11, 553–568.

[2] Esteva, A., Kuprel, B., Novoa, R.A., Ko, J., Swetter, S.M., Blau, H.M., Thrun, S., 2017. Dermatologist-level classification of skin cancer with deep neural networks. Nature 542, 115–118.

Investigating neuromorphic devices to improve sensing, vision & control in surgical robots

Supervisors: Elizabeth Rendon and Rodrigo Aviles

From Deep Blue to AlphaGo, artificial intelligence and neural networks have become an important research direction. Neuromorphic computing is a complete rethinking of computer architecture from the bottom up. The goal is to apply the latest insights from neuroscience to create integrated circuits that function like the human brain. Neuromorphic devices replicate the way neurons are organized, communicate, and learn at the hardware level. Future neuromorphic processors are defining a new model of programmable computing.

This project investigates the integration of sensory and neuromorphic devices, to meet the sensing and learning capabilities of neural systems, by which can make the systems sense external stimulus and store the relevant information [1]. Novel neuromorphic processor integrated circuits (IC) – structural design and optimisation-, will be studied to perform neurophysiological simulations based on memristors. Sensing, imaging, and control systems used in surgical robotics will be integrated [2] and run with a xx-neuron spiking network at selected frequencies over 250 kilohertz. The system will be evaluated for power consumption (energy efficiency), latency, structural design and its performance will be compared with available hardware accelerators.

Apply for a PhD in Engineering

[1] Zeng M, He Y, Zhang C, Wan Q. Neuromorphic Devices for Bionic Sensing and Perception. Front Neurosci. 2021;15:690950.

[2] Kim, W.S., Paik, J., 2021. Soft Bionic Sensors and Actuators. Advanced Intelligent Systems 3, 2100003.

Sensing-acting loops in the brain: how does sensory cortex incorporate information about actions and outcomes?

Supervisors: Dr Andre Maia Chagas, Prof Miguel Maravall

A classical view of sensory processing, which has often inspired hierarchical architectures for information processing, holds that neurons in sensory pathways up into the cerebral cortex respond to sensory features and that experience-dependent learning primarily involves refining these sensory responses. In contrast, recent discoveries from our group and others [1] have shown that learning a goal-directed task can cause neurons in “sensory” areas of the cortex to develop non-sensory responses to, e.g., the actions performed by the animal and even to whether those actions get rewarded. Thus sensory cortex is not just a “front end” filtering and passing sensory evidence to higher areas, but is actively involved in task processing. This implies that an animal does not sense the world independently of what it needs to feel to guide behaviour.

This project will explore the neuronal circuits underlying these complex responses, and how they depend on task properties. You will train mice to carry out different sensory-guided tasks altering the relationship of a sensory stimulus to potential actions. You will record and manipulate neuronal activity using state-of-the-art methods [2] and optimise your own equipment and tools to control and measure mouse behaviour. The project will emphasise the use of open hardware approaches to lab equipment development and dissemination [3]. We particularly encourage applications from students who are interested in this exciting and powerful methodological approach and have a quantitative background.

Apply for a PhD in Neuroscience

[1] Bale MR, Bitzidou M, Giusto E, Kinghorn P, Maravall M (2021) Sequence Learning Induces Selectivity to Multiple Task Parameters in Mouse Somatosensory Cortex. Current Biology

[2] Janiak FK, Bartel P, Bale MR, … Maravall M, Baden T (2019) Divergent excitation two photon microscopy for 3D random access mesoscale imaging at single cell resolution. bioRxiv

[3] Maia Chagas A (2018) Haves and have nots must find a better way: The case for open scientific hardware. PLoS Biol 16(9): e3000014.

Computational neurophenomenology – how neural mechanisms shape perceptual experience

Supervisors: Prof Anil Seth, and one or more of Dr Chris Buckley Dr Warrick Roseboom Dr Ivor Simpson Dr David Schwartzman Prof Andy Clark

Apply for a PhD in Informatics or Cognitive Science

This project explores machine learning and artificial intelligence (ML/AI) approaches to simulating properties of perceptual experience (‘phenomenology’), and connecting these properties to underlying neural mechanisms. The core idea is that perceptual experience depends on neurally-encoded predictions about the causes of sensory signals. However, current computational models of this process do not readily account for richness and range of perceptual phenomenology. I welcome proposals which bridge disciplines to advance this research. Areas of potential focus include generative models of visual ‘hallucinatory’ experience, and ML/AI algorithms incorporating aspects of ‘attention’. The project will include opportunities to test models using behavioural and neuroimaging data, and exploration of philosophical implications.

The project will suit a candidate with a strong background in computational neuroscience and/or machine learning, with knowledge of cognitive neuroscience and an interest in consciousness research.

[1] Hohwy, J., and Seth, A.K. (2020). Predictive processing as a systematic basis for identifying the neural correlates of consciousness. Philosophy and the Mind Sciences 1(2):3, doi: 10.33735/phimisci.2020.II.64

[2] Tschantz, A., Seth, A.K., and Buckley, C.L. (2020). Learning action-oriented models through active inference. PLoS Computational Biology. 16(4):e1007805

[3] Suzuki, K., Roseboom, W., Schwartzman, D.J., and Seth, A.K. (2017). The hallucination machine: A novel method for studying the phenomenology of visual hallucination. Scientific Reports 7(1):15982

Teaching deep learning machines to remember like humans do

Supervisors: Dr Viktoriia Sharmanska, Prof Thomas Nowotny

Apply for a PhD in Informatics

There is an open problem in deep learning - neural networks are predominantly amnesiac, i.e. they forget past tasks as soon as they face a new learning task [1]. The aim of this project is to seek new solutions for curing amnesiac neural networks based on knowledge about human memory. According to cognitive neuroscience, humans are good at remembering information based on emotional content [2,3], e.g. pictures of emotional character (especially excitement) influence long-term recognition memory in humans significantly more than neutral images.

The research question in this project is: can we design neural networks that mimic human long-term recognition memory? To address it we will need to a) define and assess the ‘memory’ of modern neural networks, b) devise new models inspired by the properties of human memory, and c) establish a protocol for hypothesis testing, e.g., using curriculum learning [4] from easy-to-remember to hard-to-remember tasks, and cross-dataset learning [5] from image and video datasets.

[1] Li Z, Hoiem D: Learning without forgetting (2016) ECCV. arXiv

[2] Burke A, Heuer F, Reisberg D. Remembering emotional events (1992), Memory & Cognition.

[3] Marchewka A, Wypych M, Moslehi A, et al. Arousal Rather than Basic Emotions Influence Long-Term Recognition Memory in Humans (2016) Frontiers in Behavioral Neuroscience.

[4] Pentina A, Sharmanska V, Lampert CH. Curriculum learning of multiple tasks (2015), CVPR.

[5] Sharmanska V, Quadrianto N. Learning from the Mistakes of Others: Matching Errors in Cross Dataset Learning (2016), CVPR.

Integrated information: from neural correlate to computational correlate of consciousness

Supervisors: Adam Barrett, Anil Seth, and Christopher Buckley

Integrated information and complexity theories of consciousness relate key properties of conscious experience to certain forms of information dynamics [1]. Strikingly, empirical neural markers of these information dynamics successfully index global states of consciousness, e.g., sleep stages and levels of anaesthesia. However, these theories remain short on hypotheses about the relationship between consciousness and computation, cognition, perception and/or intelligence. Meanwhile, prominent computational theories of perception, action, and cognition – such as predictive coding, Bayesian brain and the free energy principle – make few, if any, theoretical claims about what distinguishes conscious from unconscious mental states [2]. This project will build bridges between these theories, starting from either or both ends, and encompassing modelling and/or empirical studies. Potential directions include (i) applying `integrated information decomposition’ [3] to relate different modes and time-courses of information flow to different conscious contents; or (ii) examining how (conscious and unconscious) expectations modulate information dynamics during perception and action.

[1] Mediano, P.A.M., Seth, A.K., & Barrett, A.B. (2019). Measuring integrated information: Comparison of candidate measures in theory and simulation. Entropy 21, 17.

[2] Hohwy, J., and Seth, A.K. (2020). Predictive processing as a systematic basis for identifying the neural correlates of consciousness. Philosophy and the Mind Sciences 1(2), 3.

[3] Mediano, P.A.M., Rosas, F.E., Luppi, A.I., Carhart-Harris, R.L., Bor, D., Seth, A.K., Barrett, A.B. (2021). Towards an extended taxonomy of information dynamics via Integrated Information Decomposition. arXiv: 2109.13186.

Spiking machine learning

Supervisors: Thomas Nowotny and James Knight

Recently, new methods have been developed to train spiking neural networks on machine learning tasks, such as SuperSpike (feed-forward networks [1]), eProp (approximate gradient descent for recurrent networks [2]) and EventProp (exact gradient descent [3]). In this project you will use these, and other methods, in our GPU accelerated spiking neural network simulation framework GeNN [4,5] to solve challenging machine learning tasks, for instance scene understanding based on stereo event-based camera inputs.

Apply for a PhD in Informatics

[1] Friedemann Zenke, Surya Ganguli; SuperSpike: Supervised Learning in Multilayer Spiking Neural Networks. Neural Comput 2018; 30 (6): 1514–1541.

[2] Bellec, G., Scherr, F., Subramoney, A. et al. A solution to the learning dilemma for recurrent networks of spiking neurons. Nat Commun 11, 3625 (2020).

[3] Wunderlich, T.C., Pehle, C. Event-based backpropagation can compute exact gradients for spiking neural networks. Sci Rep 11, 12829 (2021).

[4] Yavuz, E., Turner, J. & Nowotny, T. GeNN: a code generation framework for accelerated brain simulations. Sci Rep 6, 18854 (2016).

[5] http://genn-team.github.io/genn/, accessed 2021-11-30

Embodied Trustworthy AI Agents

Supervisors: Dr Novi Quadrianto, Prof Andy PhilippidesProf Thomas Nowotny

Machine learning is already involved in decision-making processes that affect peoples’ lives. Efficiency can be improved, costs can be reduced, and personalization of services and products can be greatly enhanced. However, concerns are rising about how to ensure that deployment of automated systems will follow clear, useful principles and requirements of trustworthiness [1]. One example of this is algorithmic fairness in a dynamic setting [2] which has a time-varying component so is best dealt with as a situated and embodied problem. In this project, you will therefore combine our initial work on fairness, transparency, and robustness [3] with symbol grounding, embodiment and situatedness needing a physical body [4] contributing to the work of the PAL lab [5] as well as being a member of the be.AI Leverhulme centre.

Apply for a PhD in Informatics

[1] High-Level Expert Group on Artificial Intelligence, Ethics guidelines for trustworthy AI, November 2020.

[2] A. D’Amour, H. Srinivasan, J. Atwood, P. Baljekar, D. Sculley, and Y. Halpern. Fairness is not static: deeper understanding of long term fairness via simulation studies. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pages 525–534, 2020.

[3] ERC Starting grant: Bayesian Models and Algorithms for Fairness and Transparency

[4] ActiveAI - active learning and selective attention for robust, transparent and efficient AI

[5] PAL lab

Meta plasticity, functional reconfiguration and morphological processing in motor behaviours: embodied models

Supervisors: Prof Phil Husbands, Prof Andy PhilippidesDr Chris Johnson

The ways in which multiple adaptive processes, operating at different temporal and spatial scales, interact in the nervous system and the body are not well understand, but have a crucial role in the generation of behaviour. Modulatory mechanisms (including diffusing neuromodulators) play an important part in such interactions, e.g. in meta-plasticity (the plasticity of plasticity) and in the operation of reconfiguring, multi-functional networks. Complex network dynamics, including oscillatory and chaotic dynamics, can also give rise to, and arise from, such interactions. Such processes are not confined to the nervous system. Recent work has shown how information processing can be shared between the body and the nervous system.

Apply for a PhD in Informatics

[1] Shim, Y. and Husbands, P. (2019) Embodied Neuromechanical Chaos through Homeostatic Regulation, Chaos 29(3):033123

[2] Johnson, C., Philippides, A. and Husbands, P. (2016) Active Shape Discrimination with Compliant Bodies as Reservoir Computers, Artificial Life 22(2):241-268

[3] M. O’Shea, P. Husbands, A. Philippides (2015) Nitric Oxide Neuromodulation. In Dieter Jaeger and Ranu Jung (Eds), Encyclopedia of Computational Neuroscience, vol. 3, 2087-2100, New York: Springer Reference.

[4] K. Briggman and W. Kristan (2008) Multi-functional pattern generating circuits, Ann. Rev. Neurosci. 31:271-294.

Modelling spatiotemporal uncertainty in perception and navigation

Supervisors: Ivor Simpson and one or more of Andy Philippides, Chris Buckley

Real-world inferences from spatiotemporal sensors, such as photographic or neuromorphic cameras, are often afflicted by ambiguities and uncertainties, e.g. how far away/what is that object [1,2]? Accurate modelling of these uncertainties may have significant downstream effects on avoiding hazards and selecting routes in animals and robots [3,4]. This project explores AI/ML methods for describing uncertainty from spatiotemporal data. Potential research areas include: Methods for well-calibrated and robust uncertainty estimation; fusing information from multiple sensors and over time; integrating uncertainty into decision making and learning models of environments; and investigating biological plausibility and embodied deployment of any of the previous points.

Depending on student interests’, data can be acquired using a variety of sensors and platforms including neuromorphic cameras and mobile robots. There is also potential to test hypotheses using real insect data.

Apply for a PhD in Informatics

[1] Kendall, Alex, and Yarin Gal. "What uncertainties do we need in bayesian deep learning for computer vision?." NeurIPS 2017.

[2] Dorta, Garoe, et al. "Structured uncertainty prediction networks." CVPR 2018.

[3] Graham, Paul, and Andrew Philippides. "Vision for navigation: what can we learn from ants?." Arthropod Structure & Development 46.5 (2017): 718-722.

[4] Claussmann, Laurène, et al. "A review of motion planning for highway autonomous driving." IEEE Transactions on Intelligent Transportation Systems 21.5 (2019): 1826-1848.

Ecological acoustic monitoring using machine learning and information theoretic measures

Supervisors: Ivor Simpson, Alice Eldridge, and Adam Barrett

Monitoring, understanding, and predicting the integrity of our planetary biosphere is the most critical sustainability issue of our time. The emerging science of Ecoacoustics enables the exciting possibility to eavesdrop on ecosystems to assess their health. This project will investigate how cutting-edge machine learning systems can learn compact and informative representations of such data. This will include exploration of: learning strategies, architectures and inductive biases, and capabilities for ecological monitoring and prediction. Subsequent work will evaluate the potential of using information theoretic complexity measures on learned, or heuristic, data representations to measure emergent informational dynamics of the soundscape, and investigate whether this provides a plausible marker of ecosystem integrity. This project is linked to active collaborations with ecologists at Sussex and beyond, and there will be opportunities for the students to work with a variety of data.

Apply for a PhD in Informatics

Eldridge, A., Guyot, P., Moscoso, P., Johnston, A., Eyre-Walker, Y. and Peck, M., 2018. Sounding out ecoacoustic metrics: Avian species richness is predicted by acoustic indices in temperate but not tropical habitats. Ecological Indicators, 95, pp.939-952.

Sethi, S.S., Jones, N.S., Fulcher, B.D., Picinali, L., Clink, D.J., Klinck, H., Orme, C.D.L., Wrege, P.H. and Ewers, R.M., 2020. Characterizing soundscapes across diverse ecosystems using a universal acoustic feature set. Proceedings of the National Academy of Sciences, 117(29), pp.17049-17055.

Mediano, P.A., Rosas, F., Carhart-Harris, R.L., Seth, A.K. and Barrett, A.B., 2019. Beyond integrated information: A taxonomy of information dynamics phenomena. arXiv preprint arXiv:1909.02297.

Fall propensity prediction through complexity measures and wearable technologies

Supervisors: Prof. Daniel Roggen, Prof. Luc Berthouze, + clinical partner to be identified

Falls are a major problem in the elderly as well as a number of clinical populations (Parkinson's Disease for example). Falls impact both on quality of life and health, but also incur major costs to the NHS. The aim of this project is to improve risk stratification early by quantifying/sensorising current assessments through: (a) using wearable sensors (e.g., [1]) to monitor patients during daily life (i.e., in the home rather than in a clinical setting) and extract information regarding their balance; and (b) further developing our signal processing techniques (e.g., [2-3]) focused on characterising variability in postural stability and use machine learning to build a predictive model. This project would suit a student interested in research at the interface between neuroscience, mathematics and engineering.

Apply for a PhD in Informatics

[1] Oishi, N., Heimler, B., Pellatt, L., Plotnik, M., & Roggen, D. (2021, September). Detecting Freezing of Gait with Earables Trained from VR Motion Capture Data. In 2021 International Symposium on Wearable Computers (pp. 33-37).

[2] Berthouze, L., & Farmer, S. F. (2012). Adaptive time-varying detrended fluctuation analysis. Journal of neuroscience methods, 209(1):178-188.

[3] West, Farmer, Berthouze et al. (2016). The Parkinsonian subthalamic network: Measures of power, linear, and non-linear synchronization and their relationship to L-DOPA treatment and OFF state motor severity. Frontiers in Human Neuroscience 10:517.

Collective Intelligence in Animal Construction

Supervisors: Prof Paul Graham and a supervisor from Informatics

Animals demonstrate swarm intelligence in group foraging, decision making and construction. Such natural collective behaviours show many characteristics that are desirable for engineered systems such as robustness, scalability, fault tolerance and flexibility. Thus understanding natural collective behaviour is an exciting opportunity for AI and robotics, alongside the inherent biological fascination. We will examine nest construction in wood ants, whose nest mounds are large impressive structures. Unlike well-studied construction with adhesive materials, such as termite mounds built from mud or wasp nests built with wood pulp paper, wood ant nests are built from twigs litter and other loose materials. Construction with loose materials not only requires complex manipulation skills at the level of individual ants, but also collective processes to deal with stochasticity in the positions of the component parts of the nest. This multidisciplinary project will combine field observations, lab studies and modelling to provide new insights into collective construction. Applicants are encouraged if they have relevant experience in any of these areas and are keen to learn new skills.

Apply for a PhD in Biology

Making deep neural networks see like animals do

Supervisors: Ben Evans and a co-supervisor from Biology

Over the past decade, Deep neural networks have achieved highly impressive perceptual feats. In particular, convolutional neural networks have surpassed humans in their ability to classify naturalistic images. These findings, together with their similarities in architecture and activity patterns, have sparked major renewed interest in DNNs as models of the brain. However, these models diverge from human perceptual abilities in important and unexpected ways, e.g., through their weakness to noise, susceptibility to adversarial attacks and inability to recognise stylised representations. This has been shown to be due to their over-reliance on diagnostic but brittle features such as texture [1] or even a single pixel correlated with the image class [2]. Some of these shortcomings have been addressed through incorporating more of the properties of the primate visual system into these models [3] or by using more extensive and naturalistic training regimes [4]. The aim of this project is to draw further inspiration from nature to explore which additional properties of the model or environment can further align their perceptual abilities with the robustness and generalisability of human vision.

Apply for a PhD in Informatics

[1] Geirhos, R., Rubisch, P., Michaelis, C., Bethge, M., Wichmann, F. A. & Brendel, W. (2019). ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. arXiv:1811.12231.

[2] Malhotra, G., Evans, B. D. and Bowers, J. S. (2020) Hiding a plane with a pixel: examining shape-bias in CNNs and the benefit of building in biological constraints. Vision Research 174, pp. 57–68.

[3] Evans, B. D. Malhotra, G. & Bowers, J. S. (2021). Biological convolutions improve DNN robustness to noise and generalisation. bioRxiv:2021.02.18.431827.

[4] Mehrer, J., Spoerer, C. J., Jones, E. C., Kriegeskorte, N. & Kietzmann, T. C. (2021). An ecologically motivated image dataset for deep learning yields better models of human vision. Proceedings of the National Academy of Sciences, 118(8).

Mapping the origins of animal movement and motor coordination

Supervisors: Prof Claudio R. Alonso, Dr Lucia Prieto-Godino and Prof Luc Berthouze

Movement is a defining trait of animals and robotic systems, but how the neural networks in the developing animal brain manage to adopt their specific connectivities, and generate and modulate their activities to control movement, is currently unknown. In particular, we still do not know what controls the transitions between the uncoordinated motor activities in developing embryos, into the precise patterns of movement observed in fully-formed organisms. This project will exploit the simplicity and genetic accessibility of the fruit fly Drosophila melanogaster to investigate the origins of animal movement, and use computational modelling to generate testable hypotheses on circuit structure and coordination. For this we will combine modern experimental work (genetic, optogenetic, advanced microscopy, deep neural network quantitative behavioural analysis), with state-of-the-art modelling approaches (neural selection, signal processing, time-series and dynamical systems analyses). This genuinely interdisciplinary effort will thus help us define the mechanisms underlying the emergence of coordinated movement, and extract core principles for the design of effective robotic locomotion and adaptation.

Apply for a PhD in Neuroscience


[1] Picao-Osorio, J, Johnston, J., Landgraf, M., Berni, J. and Alonso, C.R. (2015) microRNA encoded behavior in Drosophila, Science 350:815-20.

[2] Issa, A.R., Picao-Osorio, J., Rito, N., Chiappe, M.E. and Alonso, C.R. (2019) A Single MicroRNA-Hox Gene Module Controls Equivalent Movements in Biomechanically Distinct Forms of Drosophila, Current Biology 29:2665-2675.e4.

[3] Prieto-Godino, L., Diegelmann, S. and Bate, M. (2012) Embryonic Origin of Olfactory Circuitry in Drosophila: Contact and Activity-Mediated Interactions Pattern Connectivity in the Antennal Lobe, PLoS Biology 10(10):e1001400.

[4] Berthouze, L. and Goldfield E.G. (2008) Assembly, tuning, and transfer of action systems in infants and robots, Infant and Child Development 17:25-42.

[5] Hartley, C, Farmer, S. and Berthouze, L. (2020) Temporal ordering of input modulates connectivity formation in a developmental neuronal network model of the cortex, PLoS One 15(1):e0226772.

[6] Loveless. J., Garner, A., Issa, A.R., Roberts, R., Webb, B., Prieto-Godino, L., Ohyama, T. and Alonso, C.R. (2020) A physical theory of larval Drosophila behaviour BioRxiv doi.org/10.1101/2020.08.25.266163.

Text Logo for be.AI

How to Apply | Apply Now | Suggested Supervisors | Project Ideas | People in be.AI