Optimal forgetting: Semantic compression of episodic memories

Abstract

It has extensively been documented that human memory exhibits a wide range of systematic distortions, which have been associated with resource constraints. Resource constraints on memory can be formalised in the normative framework of lossy compression, however traditional lossy compression algorithms result in qualitatively different distortions to those found in experiments with humans. We argue that the form of distortions is characteristic of relying on a generative model adapted to the environment for compression. We show that this semantic compression framework can provide a unifying explanation of a wide variety of memory phenomena. We harness recent advances in learning deep generative models, that yield powerful tools to approximate generative models of complex data. We use three datasets, chess games, natural text, and hand-drawn sketches, to demonstrate the effects of semantic compression on memory performance. Our model accounts for memory distortions related to domain expertise, gist-based distortions, contextual effects, and delayed recall.

A generative model of memory construction and consolidation

“… 51 models how both individual examples and statistical regularities could be learned within HF. Post-consolidation episodic memories are more prone to schema-based distortions in which semantic or contextual knowledge influences recall 6 , 52 , consistent with the behaviour of generative models 32 . Neural representations in the entorhinal cortex (EC) such as grid cells 53 are thought to encode latent structures underlying experiences 31 , 54 , and other regions of the association cortex, such as the medial prefrontal cortex (mPFC), may compress stimuli to a minimal representation 55 .…”

“…Post-consolidation episodic memories are more prone to schema-based distortions in which semantic or contextual knowledge influences recall 6 , 52 , consistent with the behaviour of generative models 32 .…”

Unique effects of sedatives, dissociatives, psychedelics, stimulants, and cannabinoids on episodic memory: A review and reanalysis of acute drug effects on recollection, familiarity, and metamemory.

“…Based on our findings, two opposing factors in how drugs impact memories from an event may influence metamemory, the amount of memory content and memory stability (Figure 12). When one has an abundance of memories from an event, they could become more confusable with each other, especially after the passage of time in which more interfering information is encoded and memories undergo postencoding transformations such as compression (Figure 12a; D’Argembeau et al, 2022; Nagy et al, 2020). Thus, impairing certain memories or memory features could better differentiate those memories from an event that survive the amnestic manipulation, resulting in a relative enhancement of what one knows about an event (Figure 12b).…”

Scaling models of visual working memory to natural images

“…We also include the latent representation from a β-Variational Autoencoder (β-VAE) [15] as a more sophisticated baseline. Deep autoencoder models have been explored as tools to learn better image and video compression algorithms for technological applications [16,17], as well as to model human visual memory [18][19][20]. In addition to baseline models, we consider networks trained on the ILSVRC ImageNet classification challenge (both the 1,000-way and 22,000-way versions) and networks trained on the Contrastive Language-Image Pre-training (CLIP) objective [21].…”