Interneuron Diversity Enhances Temporal Computation in Biofidelic RSNNs

We investigate the distinct roles of excitatory (E) and inhibitory (I) neuronal types and the computational utility of neuronal diversity. We built biologically realistic Recurrent Spiking Neural Networks with three interneuron classes: PValb, SST, and 5HT3aR. We trained our networks on variations of a temporal sine wave generation task.

After training with backpropagation through time, the distinct I classes performed roles analogous to their cortical counterparts, optimizing temporal processing. Tuned PV neurons formed strong projections to tuned E neurons, precisely regulating their activity, while 5HT3 neurons developed robust E and I connections, ensuring balanced activity. In the working memory task variant, decreased 5HT3 activity lowered overall network inhibition, resulting in a higher E-to-I ratio that sustained the lost input signal. Our model suggests that distinct I classes create varied timescales: PV neurons have immediate effects on E neurons, while SST neurons exert a slower influence. The 3-I-model, by incorporating distinct timescales, exhibits precise excitatory regulation and smooth phase transitions.

We propose that distinct properties of cortical interneuron types enable unique computational roles, a dynamic replicated in our RSNN models. Thus, RSNNs with varied interneuron classes serve as robust models for exploring neural computation mechanisms.