Implementing engrams from a machine learning perspective: the relevance of a latent space

TL;DR

The paper investigates the hypothesis that brain engrams can be implemented as autoencoder-like recurrent networks, focusing on how latent-space dimensionality $n$ relates to the intrinsic data dimension $m$ (in an ambient space of dimension $d$, with $m \le d$). It argues that the data’s intrinsic dimension informs the required latent-space size and discusses how connectome structure across species constrains feasible latent spaces, including examples from C. elegans and mammalian brains. It then outlines a biologically plausible latent-space instantiation and explores how concept neurons could bind multimodal representations to form episodic memory in the hippocampus, drawing connections to structured state-space models. The paper highlights a fundamental brain-architecture-imposed limit on cognitive capacity, contrasts this with the potentially unbounded latent spaces available to machine-learning systems, and calls for developing theoretical frameworks to exploit these expanded latent spaces in artificial systems.

Abstract

In our previous work, we proposed that engrams in the brain could be biologically implemented as autoencoders over recurrent neural networks. These autoencoders would comprise basic excitatory/inhibitory motifs, with credit assignment deriving from a simple homeostatic criterion. This brief note examines the relevance of the latent space in these autoencoders. We consider the relationship between the dimensionality of these autoencoders and the complexity of the information being encoded. We discuss how observed differences between species in their connectome could be linked to their cognitive capacities. Finally, we link this analysis with a basic but often overlooked fact: human cognition is likely limited by our own brain structure. However, this limitation does not apply to machine learning systems, and we should be aware of the need to learn how to exploit this augmented vision of the nature.

Figures (2)

• Figure 1: Architecture of a basic neuronal network autoencoder. Latent space neurons are drawn in green.
• Figure 2: Proposed layout from a pyramidal 'concept' neuron (soma as red triangle) to four pyramidal neurons (soma in green) in the latent space. For illustrative purposes, axons are drawn in orange and dendrites in green, while potential synapses are marked as blue small circles. It should be noted that the concept neuron has the possibility of multiple axon-dendrite connections towards the neurons configuring the latent space. This simplified graph does not reflect the complexity of these very dense 3D networks, nor the role of other neuron types, in particular inhibitory ones.