A study on recurrent neural networks trained with excitatory-inhibitory constraint

C. JARNE

Conferencia; 30th Annual Computational Neuroscience Meeting: CNS*2021; 2021

Organization For Computational Neurosciences

Characterizing the dynamics of recurrent neural networks trained to perform tasks similar to those performed by animals and humans in laboratory experiments is crucial to understanding which connectivity models best predict the behavior of different areas of the brain, such as the cortex, and more specifically the prefrontal cortex (Sajikumar et al., 2014 Aug 19). In the last decades, simple models of recurrent neural networks have been successfully used to explain different mechanisms such as decision-making, motor control, or working memory (Cichon & Gan, 2015 Apr). One of the aspects that are omitted generally in those models is that neurons present differences between excitatory and inhibitory units (Dale´s Law). Building recurrent networks that present this characteristic presents several challenges (Sezener et al., 2021). In present work, the different dynamical behaviours obtained when training networks with different proportions of excitatory and inhibitory units were analyzed considering decision-making tasks. The dynamical behaviour, the performance of training and different constraints were studied. The emergent properties of the system were studied by comparing them with the results obtained with models that do not distinguish between excitatory and inhibitory units. We considered the case where the amount of excitatory and inhibitory units is balanced, and also what happens when this balance is broken.