Hippocampal Construction × Self-Play
Type: intersection Slug: intersection—hippocampal-construction-self-play Parents: claim—hippocampus-as-construction-system, claim—self-play-sufficiency Last updated: 2026-05-14 Epistemic status: Extrapolative
The intersection
The hippocampus constructs novel episodic scenes by recombining stored elements (people, places, objects) from distributed cortical stores. Self-play generates novel game strategies by recombining board states through iterative play against itself. Both systems produce outputs that transcend their training distribution — generating things that have never been experienced before. Neither system relies on external supervision for the quality of its outputs: the hippocampus evaluates constructed scenes against internal constraints (plausibility, coherence), and self-play evaluates strategies against the opponent (which is itself).
What the corpus implies
The two ideas are never connected in any of the 45 sources. The hippocampal construction hypothesis (2007) is framed purely as a neuroscience claim about memory and imagination. Self-play sufficiency (2017) is framed purely as an AI training paradigm. Yet Hassabis was the lead author on both — the intellectual continuity is personal, not textual.
The closest the corpus comes is the 2017 Neuron review (paper—neuroscience-inspired-artificial-intelligence), which argues for bidirectional inspiration between neuroscience and AI but does not make this specific connection.
What’s missing
- No paper asks whether the hippocampus might implement something analogous to self-play during imagination — generating candidate scenes, evaluating them against an internal model, and refining.
- No paper asks whether self-play systems develop internal representations that resemble hippocampal scene construction.
- The “construction without supervision” aspect is discussed in the imagination papers but never formalised as a learning algorithm that could be compared to self-play.
Generative potential
Testable prediction: If hippocampal construction is computationally analogous to self-play, then hippocampal activity during imagination should show signatures of iterative refinement — early “draft” constructions followed by evaluation-driven revision. This could be tested with high-temporal-resolution fMRI during imagination tasks.
AI direction: A self-play system equipped with a construction module (recombining stored elements rather than sampling from a policy) might exhibit more sample-efficient learning, because construction constrains the search space to plausible recombinations rather than arbitrary states.
Unifying theory: Both systems may be instances of a general principle: generative systems that improve by critiquing their own outputs eventually exceed the quality of their training data. This would connect construction, self-play, and the learnable nature conjecture into a single framework.
Falsification: If construction requires hippocampal replay but self-play does not (or vice versa), the shared recombination mechanism claim is false.