Data augmentation as a framework for modeling hippocampal contributions to generalization
Tyler Bonnen, Andrew Kyle Lampinen
q-bio.NC
2026-08-02
UPenn and Anthropic researchers argue hippocampal replay, consolidation and task-dependent retrieval are all data augmentation, turning verbal memory theories into testable models.
The hippocampus is the brain structure for episodic memory; patients with damage cannot form new memories and also fail at flexible inference. Six decades of work produced a shelf of theories: cognitive maps, pattern separation, Complementary Learning Systems, predictive maps. Each explains a slice of the data, and none shares a language with the others. A more practical problem sits underneath: most computational models run on hand-designed, idealized inputs such as one-hot vectors, so the mapping between what the model computes and what the experiment manipulated has to be asserted by hand. The linking function between evidence and theory is missing.
Tyler Bonnen (UPenn) and Andrew Kyle Lampinen (Anthropic) propose a deliberately simple frame: treat data augmentation, the ML practice of transforming existing data into new training signal, as the computational description of what the hippocampus does. They organize it on two timescales:
The load-bearing claim concerns linking functions. Augmentation operations are defined on raw sensory data, so a stimulus-computable model can serve as its own linking function: competing theories become explicit, comparable modeling choices over the same inputs. Whether shortcuts are composed inside the hippocampus or emerge from hippocampal-cortical exchange, whether replay reinstates veridically or transforms, those questions stop being verbal and start being code.
No new experiments and no numbers. This is a perspective paper. Its entire argument is a reorganization of prior literature: Box 1 shows CLS, context-driven reactivation and predictive maps all reading as augmentation; lesion deficits in transitive inference are lined up next to the reversal and two-hop limits of language models.
For ML practitioners, the paper hands back a biological license for a familiar design choice: the hippocampus stores experiences in a relatively raw format for later reinterpretation instead of compressing them, which is the retrieval-versus-parameters tradeoff in biological form. For neuroscience, it offers a route from verbal theory to executable model, where previously implicit disagreements must become explicit choices.
The authors are explicit: this is not a new theory, nothing has been built, and experimental paradigms rich enough to support nontrivial generalization from naturalistic input "remain scarce". A model targeting even one behavior, say identifying a novel shortcut, does not yet exist. The argument is also analogy-driven end to end, and data augmentation itself lacks a theory of why it works, so one under-explained thing is being used to organize another.