Dr James Bennett
Neural algorithms, computational neuroscience, and NeuroAI
We work at the intersection of neuroscience and AI, in a field popularly referred to as NeuroAI. We therefore often take one of two approaches to our research: 1) starting with principled algorithms used in AI, and using computational models to investigate whether and how they might be implemented in known neural circuits in the brain, and 2) recognising phenomena and their functionality in the brain, and investigating methods to incorporate them into and to augment algorithms used in AI. This has led to several different avenues of research, described below, each of which could motivate an MSc and/or PhD project.
1. Reinforcement learning and its implementation in dopaminergic circuits of the fruit fly, Drosophila melanogaster. We’re especially interested in how the mushroom body, an important brain region for learning, might implement hierarchical learning algorithms that handle contextual information, balance the tradeoff between learning and forgetting, and support risk-sensitive decision-making. [1, 2].
2. Inhibitory plasticity for efficient coding and regularisation of neural representations.
The vast majority of artificial neural networks (ANNs) allow individual neurons to support both positive (excitatory) and negative (inhibitory) parameters and firing rates. Real brains, however, are restricted to positive firing rates, and neurotransmitters that are either excitatory or inhibitory, but not both. We’re interested in exploiting these neurobiological phenomena to augment ANNs for AI, as well as drawing parallels between techniques used in AI with inhibitory function in the brain. [3, 4].
3. Deep learning to categorise animal movements.
Our understanding of the brain is informed, to a large extent, by the behaviours that are affected when normal brain function is perturbed. We’re interested in developing an analysis pipeline that uses deep learning to automatically extract distinct behavioural features and categorise them. We ultimately wish to develop a tool to augment modern tracking software, and to make the tool easy to use for the community. [5, 6]
We collaborate with several groups in Informatics (Thomas Nowotny, Andy Philippides, Jamie Knight, Maxine Sherman) and in Neuroscience (Claudio Alonso, Paul Graham), as well as groups at other institutes.
Key references
[1] Modi et al. (2020), The Drosophila Mushroom Body: From Architecture to Algorithm in a Learning Circuit, Annu. Rev. Neurosci. 43, 465–84, DOI: 10.1146/annurev-neuro-080317-0621333
[2] Bennett et al. (2021), Learning with reinforcement prediction errors in a model of the Drosophila mushroom body, Nat. Comms. 12:2569, DOI: https://doi.org/10.1038/s41467-021-22592-4
[3] Cornford et al. (2021), Learning To Live With Dale’s Principle: ANNs with Separate Excitatory And Inhibitory Units, ICLR, DOI: https://doi.org/10.1101/2020.11.02.364968
[4] Sprekeler (2017), Functional consequences of inhibitory plasticity: homeostasis, the excitation-inhibition balance and beyond, Curr. Op. Neurobiol. 43, 198–203, DOI: https://doi.org/10.1016/j.conb.2017.03.014
[5] Mathis et al. (2018), DeepLabCut: markerless pose estimation of user-defined body parts with deep learning, Nat. Neurosci. 21, 1281–1289, DOI: https://doi.org/10.1038/s41593-018-0209-y
[6] Balestriero et al. (2023), A Cookbook of Self-Supervised Learning, arXiv:2304.12210, DOI: https://doi.org/10.48550/arXiv.2304.12210
Visit Dr James Bennett's profile for a full list of publications.