Fig. 1. Overall connectivity of the hippocampus. See text for details. (PP: perforant path; DG: dentate gyrus).

Fig. 2. The hetero-associative HLAM model. See text for explanation. I. Fast loop for completion of index patterns across modalities in CA3; II. Reentrant slow loop for pattern completion.

Fig. 3. Comparison of the storage capacity in bits per connection of the HLAM and the associative net for an input layer of size 400 and different numbers of active input nodes.

Fig. 4. Recall capacity using different sparse codes in the hidden layer. Circle: standard HLAM with randomly chosen hidden layer representation with k=10 and decorrelation procedure (as described in the text); about 140 patterns of size N=400 with L=100 can be restored with 99% accuracy. Diamond: clustered sparse representation in the hidden layer using 3×3 square regions and the decorrelation procedure using N/2 preselected units. Cross: same as diamond but only N/4 units were preselected. Rest: various strategies without decorrelation procedure.

Fig. 5. Sparse representation in the 50×50 hidden layer of the HLAM. The upper row shows the summed up activity of all 3×3 square regions, which were used to represent different inputs, after 100, 200, 300 and 400 patterns have been learned by the network. The lower row shows the restored hidden layer representations from the input. Darker gray-values correspond to the number of recruitments of each neuron in multiple sparse codes.

Fig. 6. Recall from incomplete input (10%) for the two recall thresholds k (1) and k−1 (2). The graphs show the results for a single run. (BLRN=hidden layer rogue nodes; FLRN=feature layer rogue nodes; and FLM=feature layer misses.)

Fig. 7. Recall from incomplete input (10%). The graphs show the results summed up over ten runs. (HLRN=hidden layer rogue nodes; FLRN=feature layer rogue nodes; FLM=feature layer misses.) On average, 60 patterns can be recalled with 99% accuracy from 10% of the input.

Fig. 8. Information flow in the modular hippocampus, see text for details.

Fig. 9. Recall capacity of the modularized HLAM from one module; about 125 patterns of size N=800 with L=200 can be restored with 99% accuracy.

Fig. 10. Episodic processing in region CA3: the five events (T=−5,…,T=−1) which occur at successive intervals during one theta cycle predict the event at T=0.

Fig. 11. Comparison of the storage capacity of random sequences generated from 10, 20, 30, 40, and 50 random patterns.