AI-driven adaptive learning models in neuroscience: model-data alignment and hippocampal networks

Understanding how learning and credit assignment (CA) occur in the brain remains an open question. Backpropagation and other techniques used for CA in artificial neural networks, are based on biologically unrealistic requirements, e.g., symmetrical backward and forward synapses weights [1]. Recent advances at the intersection of neuroscience and deep learning (DL), have led to the development of DL models, which incorporate known neural structures and brain mechanisms, enabling CA in a biologically realistic manner [2]. However, the advancement of these models and the bridging of the gap between them and clinical neuroscience data is necessary.
Towards this direction, we will develop a DL framework to align biologically plausible neural networks (BPNN) with neuroimaging datasets. Also, we aim to advance BPNNs by introducing mathematical modifications modelling cognitive decline in cortex models [2], and by developing a model capturing the adaptive nature of hippocampal learning, whose disruption is observed in cognitive decline [3]. We will distinguishably map BPNNs modelling cognitive decline to cases', e.g., Alzheimer's, fMRI data.
Aim 1: Develop neuroAI-clinical alignment framework.
First, we will develop a DL autoencoding framework enabling the comparison between the activation patterns of different networks, as well as, their mapping to real fMRI data. We aim to compare the activation patterns of BPNNs with the ones of other conventional models to explore if they are more meaningful, i.e., resembling more the way the brain encodes information.
Aim 2: Test neuroAI-clinical framework: align NN modelling cognitive decline with fMRI data.
Then, working upon the alignment between BPNN neural responses and real fMRI data, we will focus on modelling cognitive decline mechanisms as a modification/modulation, or the absence of it, in BPNNs. This will be validated using the DL framework to align the activation patterns of BPNNs to cases' fMRI data and the ones of the modified version of BPNNs to controls' fMRI data.
Aim 3: Develop adaptive learning model of hippocampal learning.
Finally, we will develop an adaptive learning (ADAM-like) model of hippocampal learning, i.e., introduce a mathematical framework or modulations to architectures used to model the hippocampus, describing the way the hippocampus adapts its learning processes based on experience and changing conditions, e.g., through mediation by hippocampal forward replays. This type of model would capture the dynamic nature of hippocampal learning, e.g., its ability to form, update, and retrieve memories, processes whose disruption occurs in cognitive decline [3].
All the aims above are focused on understanding and modelling learning processes in the brain and their disruption in neurological conditions, such as Alzheimer's disease. This research employs a novel methodology, as it brings different disciplines together by integrating advances in DL models, theoretical neuroscience and clinical data. The alignment of BPNNs activation patterns with clinical data will highlight their validity. The introduction of modulations in cortex networks and the modelling of hippocampal learning that will allow the distinct mapping to cases' fMRI data will be a step towards modelling cognitive decline. This project falls within the intersection of the Biological informatics and Artificial Intelligence technologies EPSRC research areas and such an approach has the enormous potential to link, for the first time, cellular neuroscience and the cognitive decline observed across a range of conditions.
[1] Lillicrap, Timothy P., et al. "Backpropagation and the brain." Nature Reviews Neuroscience 21.6 (2020)
[2] Greedy, Will, et al. "Single-phase deep learning in cortico-cortical networks" Advances in neural information processing systems 35 (2022)
[3] Pedamonti, Dabal, et al. "Hippocampal networks support reinforcement learning in partially observable environments." bioRx