LOOPER: Inferring computational algorithms enacted by neuronal population dynamics

Abstract: Recording simultaneous activity of hundreds of neurons is now possible. Existing methods can model such population activity, but do not directly reveal the computations used by the brain. We present a fully unsupervised method that models neuronal activity and reveals the computational strategy. The method constructs a topological model of neuronal dynamics consisting of interconnected loops. Transitions between loops mark computationally-salient decisions. We accurately model activation of 100s of neurons in the primate cortex during a working memory task. Dynamics of a recurrent neural network (RNN) trained on the same task are topologically identical suggesting that a similar computational strategy is used. The RNN trained on a modified dataset, however, reveals a different topology. This topology predicts specific novel stimuli that consistently elicit incorrect responses with near perfect accuracy. Thus, our methodology yields a quantitative model of neuronal activity and reveals the computational strategy used to solve the task.