Training Spiking Neural Networks in pure Julia
Ben Arthur, Christopher Kim
Training artificial neural networks to recapitulate the dynamics of biological neuronal recordings has become a prominent tool to understand computations in the brain. We present an implementation of a recursive-least squares algorithm to train units in a recurrent spiking network. Our code can reproduce the activity of 50,000 neurons of a mouse performing a decision-making task in less than an hour of training time. It can scale to a million neurons on a GPU with 80 GB of memory.