Fig 1.
Overview of model architecture and learning.
(A) The model includes major hippocampal subregions: DG, CA3, and CA1, along with the superficial and deep layers of the EC. It operates in two distinct theta-related phases: encoding and retrieval. (A1) Information flow during encoding. The superficial EC provides inputs to the DG and deep EC, but not directly to CA3. Key connections involved in CA3 (the PP from EC, Rc within CA3, and the Sc to CA1) are silenced during encoding, so synaptic plasticity occurs at those sites without triggering immediate CA3 output. (A2) Information flow during retrieval. The superficial EC sends the cue directly into CA3, which then reactivates the stored memory trace through Rc, CA1, and deep EC. (B) Symmetric STDP kernel used during learning. (C) Summary of synaptic plasticity across connections in the model. Abbreviations: ext: external input; mf: mossy fiber; exc: excitatory neuron; inh: inhibitory neuron.
Fig 2.
Illustrations of CA3 engram formation inducing selective inhibition.
(A) The engram formation in CA3 driven by DG induces selective inhibition. Inhibitory neurons activated by mossy fibers inhibit most of the excitatory neurons. The surviving excitatory neurons and the activated inhibitory neurons form a neural assembly through STDP. (B) Distribution of CA3 neurons based on the number of stored memories. (C) Characterization of CA3 connections with inhibitory neurons for global versus selective inhibition. In the case of global inhibition, inhibitory neurons suppress excitatory neurons without regard to which engram they represent. In contrast, for selective inhibition, inhibitory neurons associated with a specific engram do not inhibit excitatory neurons within that same engram. (D) The inhibitory influence of one engram on others as a function of the degree of overlap between engrams.
Fig 3.
Structure of the dentate gyrus in the model (A)
The DG is modeled as two feedforward layers: the hilus layer and the GCL. Each layer contains excitatory (gray) and inhibitory (red) neurons. Superficial EC inputs drive all DG neurons. Inhibitory neurons in the GCL also receive inputs from excitatory neurons in the hilus layer, which are regulated by hilus inhibitory neurons. Finally, excitatory neurons in the GCL, regulated by inhibitory neurons within the GCL, send mossy fiber outputs to CA3.
Fig 4.
DG activity and pattern separation.
(A) Spike raster during encoding of 10 distinct inputs (A to J). Plotted are neurons in the superficial EC, and excitatory and inhibitory neurons in both DG layers (hilus and GCL). 50 of the 100 hilus excitatory neurons and 50 of the 400 GCL inhibitory neurons are shown for visualization. The 16 excitatory GCL neurons encoding the 10 inputs (out of 800) are highlighted. (B) Pattern separation performance of the DG. (C) Histogram of engram overlaps: the similarity of DG output (GCL excitatory) engrams versus similarity of superficial EC input patterns. Most outputs fall into low-overlap bins, indicating robust separation.
Fig 5.
Encoding of inputs and learned synapses.
(A) Weight matrices after STDP learning. Weights represent the connection weights multiplied by the learned peak conductances. Only CA3 neurons encoding the 10 example inputs are visualized; E, excitatory; I, inhibitory. (A1) Learned direct PP weight matrix from superficial EC to CA3 excitatory. (A2) Learned excitatory Rc weight matrix from CA3 excitatory to CA3 excitatory. (A3) Learned Rc weight matrix from CA3 excitatory to CA3 inhibitory. (A4) Learned Rc weight matrix from CA3 inhibitory to CA3 excitatory. (A5) Learned Sc weight matrix from CA3 excitatory to CA1. (B) Raster plot during the encoding of 10 different inputs (A to J). Shown are spikes in the superficial EC, excitatory neurons in the GCL, CA3 excitatory and inhibitory neurons, CA1, and deep EC. Other types of neurons in the DG are omitted. As in (A), only the CA3 excitatory neurons active in these examples are selected for visualization. (C) CA3 memory capacity: number of engrams formed versus number of input patterns. (D) Relationship between input pattern size and resulting CA3 engram size (points) with a linear fit.
Fig 6.
Retrieval of learned memories.
(A) Spike raster during retrieval in response to a cue similar to inputs E and G out of the 10 encoded patterns A to J. Layer conventions are the same as in Fig 3B. The cue is presented in 10 trials; some yield successful recall of E or G (indicated by strong firing in deep EC for those patterns), while others fail. (B) CA3 engram firing rates during retrieval for the example trials. (B1) Successful retrieval of pattern G: the CA3 G-engrams dominate firing. (B2) Successful retrieval of pattern E: the CA3 E-engrams dominate. (B3) Retrieval failure: no single engram dominates, and activity remains low and diffuse.
Fig 7.
Impact of the direct PP on retrieval.
(A) Illustration of different percentages of partial cues for a learned input A. (B) Firing rate heatmap for CA3 neurons in response to each cue during retrieval (120 ms). Each heatmap shows the most common firing pattern across 20 trials for that cue. (B1) With the direct PP intact, firing patterns are consistent across cue variations. (B2) With the direct path silenced, firing patterns vary strongly with each cue. (C) Retrieval performance versus percentage of cue under two conditions: with direct PP (blue) and without direct PP (red). Each point is the mean success rate (averaged over 5 random cues 20 repeats; 100% cue uses one sample).
Fig 8.
Effects of input strength disparity and engram overlap on retrieval.
(A) to (D) Input strength disparity task: (A) Cue schematic: two CA3 engrams are cued with different input strengths. (B) Example of the learned Rc weight matrix. (C) CA3 engram firing rates during retrieval; two example trials with the same cue are shown: (C1) memory A is retrieved, (C2) memory B is retrieved. (D) Retrieval performance versus input strength disparity, comparing with (blue) and without (red) selective inhibition. Performance is measured as the mean retrieval success rate over 15 bias samples 50 repeats each. (E) to (H) Overlap task: (E) Cue schematic: two overlapping engrams are cued equally. (F) Example learned Rc weight matrix. (G) CA3 engram firing rates for example trials with the same cue: (G1) memory A is retrieved, (G2) memory B is retrieved. (H) Retrieval performance versus percentage overlap between engrams, comparing with (blue) and without (red) selective inhibition.
Fig 9.
Effect of the number of competing engrams on retrieval.
(A) SCue schematic: multiple CA3 engrams are presented with the same cue. (B) Example learned Rc weight matrix. (C) CA3 engram firing rates for example trials with the same cue: (C1) memory A is retrieved, (C2) memory B is retrieved, (C3) C (C4) D, (C5) E, (C6) F. (D) Retrieval performance versus number of engrams, comparing conditions with (blue) and without (red) selective inhibition.
Fig 10.
Effect of engram size in CA3 on retrieval.
(A) Cue schematic: two CA3 engrams of different sizes are cued. (B) Example learned Rc weight matrix. (C) CA3 engram firing rates for example trials with the same cue: (C1) memory A is retrieved, (C2) memory B is retrieved. (D) Retrieval performance versus engram size differences, comparing conditions with (blue) and without (red) selective inhibition. (E) Dominance of the larger engram: fraction of successful trials retrieving the larger-memory engram, versus size differences, comparing with (blue) and without (red) selective inhibition.
Fig 11.
Pattern separation task and results.
(A) Task schematic: Input A is presented in one 120-ms encoding phase. Input B is presented over 7.2 s (30 theta cycles spanning encoding and retrieval). B’s similarity to A is varied across 105 trials. (B) Example learned Rc weight matrix for (B1) highly similar ( Input = 0.92) and (B2) less similar (
Input = 0.17) inputs. (C) Change in neural representation (
output) in each subregion versus input discrimination (
input), averaged over 30 retrieval phases. (D) Same
output vs.
input during encoding only. (E) Same during retrieval only.
Fig 12.
Pattern completion task and results.
(A) Pattern completion task and results. Two orthogonal inputs (A and B, overlap = 0) are presented simultaneously during the encoding. During the retrieval, 100 cues are constructed with linearly varying similarity to A and B (overlaps sum to 1), each presented for 7.2 s (30 phases) without further learning. (B) Learned Rc weight matrix after training on A and B. (C) Retrieval performance versus cue similarity to A ( A–Cue). Performance is measured by the probability of retrieving A (blue) or B (purple) across 30 phases. (D) Change in neural representation (
output) versus cue discrimination ((
A–Cue), averaged over 30 phases.
Table 1.
Parameters for neurons organizing each layer.
Table 2.
Parameters for connections between layers.
Table 3.
Parameters of Izhikevich model for excitatory neurons and inhibitory neurons.