Last updated: 2025-11-02
CSCGs provide a mechanistic explanation for diverse hippocampal phenomena, supporting the hypothesis that space is a latent sequence: spatial representation emerges from latent higher-order sequence learning.
Clones respond to sequential contexts, not locations directly
- Place fields = overlay of clone activations onto experimenter's spatial map
- Agent doesn't decode location; uses clones directly for planning via message passing
Same location, different sequential contexts activate different clones
- T-junction: left-turn context vs right-turn context
- Natural consequence of context-dependent representation
Three types explained by single mechanism:
- Global remapping: New environment → different clone set activated
- Rate remapping: Same clones, different activation magnitudes (minor context changes)
- Rotational remapping: Contexts anchored to visual cues rotate with cues
Lap number becomes part of sequential context
- CSCG learns to distinguish laps, predict reward timing
- Event-specific remapping explained without special mechanisms
Landmarks anchor sequential contexts
- Move landmark → contexts move relative to it → place field components follow
Place fields unchanged when connectivity changes:
- Blocked doors don't change visual cues → place fields unchanged
- Planning (replay messages) adapts by updating transition graph
- Behavior changes without place field remapping
Multiple neurons per observation
- Lateral connections implement transition matrix (CA3 recurrence)
- Bottom-up input implements emission matrix (entorhinal cortex → CA3/CA1)
Forward-backward as recurrent activity
- Biologically plausible via integrate-and-fire neurons (Rao 2004)
Expectation-maximization analogous to spike-timing dependent plasticity
- Local update rules, no global error signal needed (Nessler et al. 2009, 2013)
Hippocampal replay implements message propagation
- Forward replay: forward messages for prediction
- Backward replay: backward messages for goal-directed planning
Recent evidence shows grid cells not necessary for place cells (Brandon et al. 2014)
- CSCG learns spatial structure without grid input
- Grid cells could serve as optional scaffold for faster learning in large environments
The "place field methodology" (overlaying sequential neural responses onto spatial maps) may itself be a source of anomalies. True hippocampal representation: sequential contexts in latent clone space. Spatial, temporal, and abstract relational structures unified under sequence learning framework.
For detailed mathematical explanations of dozen+ hippocampal phenomena, see:
22-12_Raju_Space_is_latent.md- Unified theory with experiments21-04_George_Clone-structured_graph_representations.md- CSCG mechanisms
This summary connects CSCGs to neuroscience, showing how latent higher-order sequence learning unifies diverse hippocampal phenomena under a single computational framework.