Construction×Self-Play + Replay×Hippocampal-Replay
Type: intersection (second-order) Slug: intersection—construction-self-play-replay-consolidation Parents: intersection—hippocampal-construction-self-play, intersection—experience-replay-hippocampal-replay Last updated: 2026-05-14 Epistemic status: Extrapolative
The combination
If hippocampal construction is analogous to self-play (generating novel outputs by recombination), and experience replay is analogous to hippocampal SWR replay (consolidating by re-sampling), then the brain’s full imagination cycle is: construct during wake (self-play) → replay during sleep (consolidation) → improved construction next day. No AI system implements both phases of this cycle.
What emerges
The two first-order intersections each identify a parallel in isolation. Combined, they reveal a two-phase architecture that neither alone implies:
- Phase 1 (wake): The hippocampus runs internal self-play — generating candidate scene constructions, evaluating them against internal models, discarding bad ones.
- Phase 2 (sleep): The successful constructions are replayed to the neocortex, where they gradually become permanent semantic representations.
This maps directly onto the complementary learning systems theory (CLS): fast hippocampal learning + slow neocortical consolidation. But CLS frames this as “memory transfer,” not “self-play with consolidation.” The self-play framing adds something CLS lacks: the hippocampus doesn’t just store experiences — it generates novel ones and selects the best.
Gap
No AI system combines generative self-play with selective replay-based consolidation. Current self-play systems (AlphaGo, AlphaZero) train continuously without a sleep phase. Current replay systems (DQN) replay past experiences without generating novel ones. The brain does both.
Generative potential
Architecture: A self-play system that generates candidate solutions during “wake” phases, evaluates them, stores the best in a replay buffer, and then trains the main network from the replay buffer during “sleep” phases. This decouples generation from learning — the generator can be more exploratory because bad candidates are filtered during consolidation.
Neuroscience prediction: Hippocampal activity during REM sleep (when construction-like imagery is most vivid) should preferentially replay novel constructions rather than veridical experiences. This would distinguish the self-play-with-consolidation model from standard CLS (which predicts replay of actual experiences).
Falsification: If a single system can both construct novel scenarios and consolidate them without separate wake/sleep phases, the phase-separation requirement is false.