Quantitative investigation of memory recall performance of a computational microcircuit model of the hippocampus

Abstract

Memory, the process of encoding, storing, and maintaining information over time to influence future actions, is very important in our lives. Losing it, it comes with a great cost. Deciphering the biophysical mechanisms leading to recall improvement should thus be of outmost importance. In this study, we embarked on the quest to improve computationally the recall performance of a bio-inspired microcircuit model of the mammalian hippocampus, a brain region responsible for the storage and recall of short-term declarative memories. The model consisted of excitatory and inhibitory cells. The cell properties followed closely what is currently known from the experimental neurosciences. Cells' firing was timed to a theta oscillation paced by two distinct neuronal populations exhibiting highly regular bursting activity, one tightly coupled to the trough and the other to the peak of theta. An excitatory input provided to excitatory cells context and timing information for retrieval of previously stored memory patterns. Inhibition to excitatory cells acted as a non-specific global threshold machine that removed spurious activity during recall. To systematically evaluate the model's recall performance against stored patterns, pattern overlap, network size, and active cells per pattern, we selectively modulated feedforward and feedback excitatory and inhibitory pathways targeting specific excitatory and inhibitory cells. Of the different model variations (modulated pathways) tested, 'model 1' recall quality was excellent across all conditions. 'Model 2' recall was the worst. The number of 'active cells' representing a memory pattern was the determining factor in improving the model's recall performance regardless of the number of stored patterns and overlap between them. As 'active cells per pattern' decreased, the model's memory capacity increased, interference effects between stored patterns decreased, and recall quality improved.

Keywords: Bistratified cell; Computer model; Dendrite; Excitation; Inhibition; Medial septum; Memory retrieval; OLM cell; Pyramidal cell; Theta rhythm.

Fig. 1

Hippocampal CA1 microcircuit showing major cell types and their connectivity. SLM: Stratum lacunosum-moleculare; SR: stratum radiatum; SP Stratum pyramidale; SO: stratum oriens; PC: pyramidal cell; AAC: axo-axonic cell; BC: basket cell; BSC: bistratified cell; CA3: CA3 Schaffer collateral input; MS: medial septum. Black lines: excitatory input; blue lines: inhibitory input; maroon lines: MS inhibitory input

Fig. 2

(Left) Recall microcircuit model of region CA1 of the hippocampus and (right) CA1-PC model with one excitatory (CA3) and six inhibitory (BSC) synaptic contacts on its dendrites. EC: entorhinal cortical input; CA3: Schaffer collateral input; AAC: axo-axonic cell; BC: basket cell; BSC: bistratified cell; OLM: oriens lacunosum-moleculare cell; SLM: stratum lacunosum-moleculare; SR: stratum radiatum; SP: stratum pyramidale; SO: stratum oriens. During recall, only PCs, BSC, and OLM cell are active. AAC and BCs are inactive due to strong medial septum inhibition. BSC and PC are driven on their SR dendrites by a strong CA3 excitatory input, which represented the contextual information. EC input is disconnected from the network, which thus has no effect on it. Red circles on PC dendrites represent loaded synapses, whereas black circles on PC dendrites represent unloaded synapses

Fig. 3

Voltage traces of model cells with respect to a single theta cycle

Fig. 4

a Raster plot showing septal (top 20) and CA3 input (bottom 100) spikes. b Raster plot showing 20 ‘active cells’ activity coding for a particular memory pattern. c Twenty ‘active cells’ spike count in a sliding 10 ms bin. d Recall quality in a sliding 10 ms bin

Fig. 5

Exemplar set of five memory patterns with 40% overlap between them

Fig. 6

Schematic drawing of a ‘model 1’, b ‘model 2’, and c ‘model 3’. In ‘model 1’, a strong excitatory CA3 input increases BSC firing response, which generates on PC dendrite numerous small amplitude IPSPs, thus producing a very strong inhibitory environment which filters out spurious neuronal activities. In ‘model 2’ a strong BSC inhibitory drive to PC dendrite causes postsynaptically fewer, but with larger amplitude IPSPs. In ‘model 3’, a strong excitatory PC feedback signal to BSC increases its firing response, which generates fewer than ‘model 1’ IPSPs on PC dendrite, and hence a less strong inhibitory environment than ‘model 1’

Fig. 7

Mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of percent overlap (0%, 10%, 20%, 40%). Each model was a network of 100 PCs, 1 BSC and 1 OLM cell

Fig. 8

Mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of percent overlap (0%, 10%, 20%, 40%). Each model was a network of 300 PCs, 1 BSC, and 1 OLM cell

Fig. 9

Comparison of mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of network size (100 PCs vs 300 PCs) for different numbers of stored patterns, active cells, and 40% pattern overlap

Fig. 10

Mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of active cells per pattern (5, 10, 20) when five patterns were stored with various percentages of pattern overlap. Each model was a network of 100PCs, 1 BSC, and 1 OLM cell

Fig. 11

Mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of active cells per pattern (5, 10, 20) when ten patterns were stored with various percentages of pattern overlap. Each model was a network of 100PCs, one BSC,, and one OLM cell

Fig. 12

Mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of active cells per pattern (5, 10, 20) when five patterns were stored with various percentages of pattern overlap. Each model was a network of 300PCs, one BSC, and one OLM cell

Fig. 13

Mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of active cells per pattern (5, 10, 20) when 10 patterns were stored with various percentages of pattern overlap. Each model was a network of 300 PCs, one BSC, and one OLM cell

Fig. 14

Comparison of mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of network size (100 PCs vs 300 PCs) for five stored patterns, 20 active cells, and different % pattern overlaps

Fig. 15

Comparison of mean recall quality of ‘model 1’, ‘model 2’, and ‘model 3’ as a function of network size (100 PCs vs 300 PCs) for ten stored patterns, ten active cells, and different % pattern overlaps