Dynamic control of inhibition improves performance of a hippocampal model

Adding synaptic modification to the inhibitory interneuron in a minimal com putational model of hippocampal region CA3 improves average performance of the simulations. After training on two partially overlapping sequences, sim ulations are tested on a sequence completion problem that can only be solve d by using context dependent information. Simulations with dynamic autonomo usly scaling (DAS) inhibition are more robust than those without. In the DA S model, scaling factors for inhibition are adjusted gradually over time to compensate for the original model's tendency to move away from a pre-set a ctivity level. This variable inhibition modifies more slowly than the local , associative synaptic modification of the excitatory synapses. As a result , activity fluctuations from one time-step to the next continue to occur, b ut average activity levels show small variability across training. These re sults suggest that restricting long term activity fluctuations can be benef icial to recurrent networks that must learn context dependent sequences. (C ) 2001 Elsevier Science B.V. All rights reserved.