Shirin Dora

Training recurrent neural networks with inhibitory neurons in multisensory integration tasks.

In recent years, Recurrent Neural Networks (RNNs) trained to perform cognitive tasks have become a useful tool in neuroscience. However, many RNN implementations are lacking in terms of biophysical realism. Most notably, typical RNNs don't distinguish between excitatory and inhibitory neurons, a realistic feature of neural circuits known as Dale's rule. This limits their applicability to understand perception and cognition. In this talk, we will present our approach for using RNNs, with explicit excitatory and inhibitory units, that are trained to perform several multisensory perception tasks. We will analyze the role of excitatory and inhibitory neurons in the underlying neural computations. Our models are able to reproduce the behavioral performance of rodents in several tasks that involve detection, discrimination and integration of multisensory information. In addition, our results suggest that not just excitatory, but also inhibitory cells, develop a marked specificity to sensory and choice modalities, in agreement with recent experimental data involving multisensory stimuli. These results highlight the importance of implementing biophysically realistic constraints in RNNs, in particular selectivity in excitatory and inhibitory units, to decipher the neural computations for cognitive functions.