Modelling the formation of new memories in the human brain

The importance of learning and memory in our everyday lives is highlighted by neurological conditions such as Alzheimer's disease, whose early symptoms include forgetting about recent events, names and faces. Recent experimental work involving single neuron recordings from epilepsy patients implanted with depth electrodes for clinical reasons has shown that individual neurons in the human brain change their firing to link associations when a new memory is formed (Ison, Quiroga & Fried, Neuron 2015). Moreover, DeFalco, Ison, Fried & Quiroga (Nature Communications, 2016) recently reported that neurons in the medial temporal lobe (MTL) are involved in long-term representation of associations between items. These findings, together with recent evidence from work on rodents and non-human primates, suggest that memories are encoded in sparse assemblies of neurons. However, conventional memory models predict that the MTL is only involved in the initial representation of memories (e.g. Norman & O'Reilly, Psychol. Rev. 2003). Computational models can shed light into this conundrum, leading to a better understanding of the way in which the brain represents memories and potentially leading to novel ideas about how memories deteriorate in patients suffering from neurological disorders.

The project aims at developing neural network models with data-constrained rules for spike-timing dependent plasticity. The first stage of the project will prioritise the training of the candidate in the area of neuroscience (with emphasis on memory), in building neural models of increasing complexity, and on familiarising with machine learning techniques. The second stage will be focused on simulating the formation of associations between patterns (Objective 1). In a subsequent stage, the model will be extended to incorporate a consolidation phase to give account for long-term associations in memory (Objective 2).

The research methodology, including new knowledge or techniques in engineering and physical sciences that will be investigated

The network model will be a randomly connected recurrent neural network consisting of excitatory and inhibitory neurons. Most of the parameters, including synaptic connections for short-term synaptic plasticity, will be drawn from experimental data (see e.g. Pokorny, Ison, Rao, Legenstein, Papadimitriou & Maass, Cer Cortex 2019). Network simulations will be conducted in Python, using either the NEST Simulator (Gewaltig & Diesman, 2007) or Brian (Stimber, Brette & Goodman, 2019). Data analyses will be performed in Python or MATLAB. The possibility of extending the size of the networks to simulate will be evaluated and potentially implemented using GPUs.

Due to its interdisciplinary nature, the project is relevant both to EPSRC Research Themes (Healthcare Technologies, Mathematical sciences) as well as other Research Council priority areas (e.g. BBSRC Healthy Ageing). Importantly, it fits extremely well with the vision of the EPSRC to healthcare applications by building critical mass around UK research strengths in computational and mathematical sciences.

Key collaborators are Prof. Stephen Coombes (School of Maths, University of Nottingham) and Prof. Wolfgang Maass (Institute of Theoretical Computer Science, TU Graz, Austria). Other potential collaborators include Prof. Gabriel Kreiman (Harvard Medical School) and Prof. Itzhak Fried (UCLA Medical Center).