Analogy between AMICL/residual attention streams and hippocampal skip connections

Establish whether the Associative Memory for In-Context Learning (AMICL) model and the residual attention value stream architecture in Transformers are analogous to skip connections observed in the hippocampus, such as direct CA3-to-CA1 pathways that bypass CA2.

Background

The paper introduces AMICL and a residual attention value stream that improves in-context learning. Drawing parallels to hippocampal anatomy, the authors discuss skip-like pathways (e.g., CA3 to CA1) and speculate about potential analogies between their architectures and biological skip connections.

They explicitly state that determining this analogy remains an unresolved issue, suggesting future comparative analyses between engineered residual streams and hippocampal connectivity patterns.

References

Whether our AMICL model or residual attention stream modification can be considered analogous to skip connections witnessed in the hippocampus remains to be studied.

— Associative memory inspires improvements for in-context learning using a novel attention residual stream architecture  (2412.15113 - Burns et al., 2024) in Section Discussion

We conjecture that the explicit local features serve a role analogous to a residual connection, providing a direct gradient pathway for local coupling information.

— Universal sampling of spin systems across quenched disorder  (2609.11336 - Liu et al., 10 Sep 2026) in Supplemental Material, Section I.C, subsection “Decoder”