Biologically-Inspired Deep Predictive Learning for Episodic Memory Event Segmentation

Abstract

Computational modeling for event segmentation in the literature has been directed towards detailing individual regions of the brain, mainly the hippocampus, prefrontal cortex and substantia nigra and the role they play in episodic memory formation. Modern predictive recurrent neural networks, which have been shown to perform well on naturalistic video frame prediction, have the potential to explain critical properties of neuronal responses and perception during event segmentation. However, they have not been examined in this light. This thesis explores the ability of state-of-the-art deep predictive learning approaches to explain properties of event segmentation. We identify similar artificial neural responses in our model and biological neural responses from Intracranial Electroencephalography data. These responses are compared in the context of detecting separation between different events in naturalistic video data input. Finding such parallels could provide insight for improving the biological plausibility of deep learning networks. Furthermore, such computational models can serve as a biological proxy and a testing ground for mechanistic hypotheses that bridge neural computation to observable behavior.