Learning distributed representations with complementary components - A generative approximation framework for episodic memory encoding

Abstract

Episodic memory must preserve detail while storing memories efficiently. We propose a framework of memory encoding in which generative neural network models approximate input through their frozen weights, yielding starting points from which the same weights can reinstate the input. This removes the need for mutually invertible cortical-hippocampal mappings, realizes a tradeoff between computational effort and mnemonic precision, and uses statistical regularities of the stimulus space to find representations that are distributed across the network. We implement this principle with a toy model and with large image-generation AI models, operationalizing mnemonic precision as closer approximation. Our model reproduces key findings from the Mnemonic Similarity Task: shallow encoding selectively impairs lure rejection, as observed in aging populations. Encoding further benefits from expertise: domain-finetuned models need fewer approximation steps for images from their own domain than for images from other domains. Finally, the representational geometry of latent starting points resembles medial temporal lobe but reinstatement through the weights of the network traverses the visual hierarchy in reverse order. Our framework offers a parsimonious and flexible account of memory encoding, where a generative search for a reinstatement-capable starting point finds a path for retrieval directly.