Transitory memory retrieval in a biologically plausible neural network model

Abstract

A number of memory models have been proposed. These all have the basic structure that excitatory neurons are reciprocally connected by recurrent connections together with the connections with inhibitory neurons, which yields associative memory (i.e., pattern completion) and successive retrieval of memory. In most of the models, a simple mathematical model for a neuron in the form of a discrete map is adopted. It has not, however, been clarified whether behaviors like associative memory and successive retrieval of memory appear when a biologically plausible neuron model is used. In this paper, we propose a network model for associative memory and successive retrieval of memory based on Pinsky-Rinzel neurons. The state of pattern completion in associative memory can be observed with an appropriate balance of excitatory and inhibitory connection strengths. Increasing of the connection strength of inhibitory interneurons changes the state of memory retrieval from associative memory to successive retrieval of memory. We investigate this transition.

Keywords: Associative memory; Recurrent network; Successive retrieval of memory; Transitory dynamics.

Fig. 1

A schematic diagram of the network architecture. The network architecture consists of a population of pyramidal cells which are reciprocally connected by recurrent connections and GABAergic fast-spiking interneurons with external input and noise. Connection matrices , , and are shown in three columns at the bottom of the figure (see Appendix 3 for details)

Fig. 2

Network response for external stimulus. a Rasterplot of external stimulus. An external stimulus that was slightly different to the stored pattern was applied to each cell assembly at times 0, 250, and 500 ms, and each stimulus lasted for 40 ms. b Rasterplot of somatic spikes. These spikes are induced by the external stimulus inputs. The different colors indicate a different cell assembly. c Overlapping between a stored pattern and an activity pattern of the network. Overlapping was calculated using a time window of 10 ms, and the time window for moving average was 0.05 ms. The parameters in this simulation are g_GABA = 0.037 and g_AMPA,rec = 0.036. The simulation suggests that an attractor dynamics emerges and employs auto-association for the recall of memories

Fig. 3

The structure of the basin of attraction. a The difference between an input pattern and a stored pattern (DBIS). Plus or minus sign of DBIS indicates the increase or the decrease of the number of active neurons in the input pattern from a stored memory pattern. b The abscissa denotes DBIS, and the ordinate denotes the probability of memory retrieval. Red solid and blue dashed curves show the cases of g_GABA = 0.030 and g_GABA = 0.040, respectively

Fig. 4

Rasterplot of the somatic spikes for different values of the connection strength g_GABA. We observed the behavior of most cells firing, as in an epileptic seizure, for g_GABA = 0 (state I), associative memory retrieval for g_GABA = 0.032 (state II), and successive retrieval of memories for g_GABA = 0.048 (state III). a Time-series of the connection strength g_GABA. The connection strength starts at g_GABA = 0, and changes to g_GABA = 0.032 at 1,000 ms, and to g_GABA = 0.048 at 2,500 ms. b Rasterplot of somatic spikes. Different colors indicate different cell assemblies. c Overlapping between somatic spikes and stored memory patterns

Fig. 5

The parameter space of g_AMPA,rec/g_GABA. a The difference between the input pattern and one of the stored patterns (). The white color scale indicates the probability of memory retrieval. b The phase diagram of the dynamic states of network: an epileptic seizure (state I), a pattern completion (state II), and a successive retrieval of memories (state III). State IV is a state which could not be classified in any of the above three states