@phdthesis{discovery10193105,
          school = {UCL (University College London)},
            note = {Copyright {\copyright} The Author 2024.  Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/).  Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms.  Access may initially be restricted at the author's request.},
            year = {2024},
           title = {Hippocampus-Inspired Representation Learning for Artificial Agents},
           month = {June},
        abstract = {Spatial representations found in the hippocampal formation of freely moving mammals, such as those of grid cells, appear optimal for spatial navigation, and also afford flexible and generalisable non-spatial behaviours. In this thesis, I propose models for learning and representing the structure underlying high-dimensional observation space in artificial agents, drawing inspiration from hippocampal neuroscience.

In the first part of the thesis, I study the construction and identification of latent representations. I propose a novel model for grid cell firing based on Fourier analysis of
translation-invariant transition dynamics. I show that effects of arbitrary actions can be predicted using a single neural representation and action-dependent weight modulation, and how this model unifies existing models of grid cells based on predictive planning, continuous attractors, and oscillatory interference. Next, I consider the problem of unsupervised learning of the structured latent manifold underlying population neuronal spiking, such that interdependent behavioural variables can be accurately decoded. I propose a novel amortised inference framework such that the recognition networks explicitly parametrise the posterior latent dependency structure, relaxing the full-factorisation assumption.

In the second part, I propose representation learning methods inspired by neuroscience and study their application in reinforcement learning. Inspired by the observation of hippocampal "replay" in both temporally forward and backward directions, I show that incorporating temporally backward predictive reconstruction self-supervision into training world models leads to better sample efficiency and stronger generalisability on continuous control tasks. I then
propose a novel intrinsic exploration framework under a similar premise, where the intrinsic novelty bonus is constructed based on both prospective and retrospective information. The resulting agents exhibit higher exploration efficiency and ethologically plausible exploration strategies.

I conclude by discussing the general implications of learning and utilisation of latent structures in both artificial and biological intelligence, and potential applications of neural-inspired representation learning beyond reinforcement learning.},
             url = {https://discovery.ucl.ac.uk/id/eprint/10193105/},
          author = {Yu, Changmin}
}