Hippocampal Pattern Separation & Completion (2D)

Encoding (Dentate Gyrus)

Input pattern overlap (EC-A vs EC-B) 70%

DG sparsity (active granule cells) 8%

Recall (CA3 attractor)

Cue corruption 35%

recalling pattern A…

Live Readouts

EC input overlap

–

DG code overlap

–

CA3 recall overlap

–

Sweeps run

Attractor state

settling…

How it works

This models the classic Marr / O'Reilly–McClelland account of hippocampal memory: the entorhinal cortex (EC) sends the same input to two subregions with opposite jobs.

Pattern separation (dentate gyrus):
  drive_i = Σ_j W_ecdg[i,j]·x_j
  y_i = 1 if drive_i is among the top k·N_dg
        0 otherwise      (sparse k-WTA code)

Pattern completion (CA3, Hopfield-style):
  W_ij = (1/N) Σ_p s_i^p s_j^p     (Hebbian storage, i≠j)
  s_i(t+1) = sign(Σ_j W_ij s_j(t))  (async recurrent update)

DG is a large, sparse, expansive layer with strong lateral inhibition (k-winners-take-all). Even two EC inputs that overlap 70–90% get recoded into nearly orthogonal sparse DG codes — this is pattern separation, and it is what stops similar memories (where you parked yesterday vs. today) from interfering with each other.
CA3's dense recurrent collaterals store patterns as attractor states of a Hebbian associative network. Starting from a corrupted or partial cue, asynchronous updates pull the network back to the nearest stored attractor — pattern completion, the basis of recalling a whole episodic memory from a fragment or a smell.
Overlap slider sets how similar the two EC input patterns are. Sparsity slider sets k, the fraction of DG cells allowed to fire — lower k separates patterns harder but stores less detail. Cue corruption sets how many CA3 bits are flipped before recall starts.
DG code overlap is measured as the Jaccard overlap of only the active granule cells in each code (|A∩B| / |A∪B|) rather than a raw bit-match fraction — with only ~8% of cells firing, a raw match count is swamped by cells that are silently off in both patterns and would misleadingly read as "highly similar" even when separation is working well.

Real damage to this circuit (hippocampal sclerosis, early Alzheimer's) degrades exactly these two functions — patients confuse similar memories (failed separation) or cannot retrieve memories from partial cues (failed completion).

EC — cortical input DG — sparse separated code CA3 — recalled attractor

Drag to pan · scroll/pinch to zoom