FixedPointFinder: A Tensorflow toolbox for identifying and characterizing fixed points in recurrent neural networks

Recurrent neural networks (RNNs) are powerful function approximators that can be designed or trained to solve a variety of computational tasks. Such tasks require the transformation of a set of time-varying input signals into a set of time-varying output signals. Neuroscientists are increasingly interested in using RNNs to explain complex relationships present in recorded neural activity (Pandarinath et al., 2018) and to propose dynamical mechanisms through which a population of neurons might implement a computation (Mante, Sussillo, Shenoy, & Newsome, 2013; Remington, Narain, Hosseini, & Jazayeri, 2018). Once fit to neural recordings or trained to solve a task of interest, an RNN can be reverse-engineered to understand how a computation is implemented in a high-dimensional recurrent neural system, which can suggest hypotheses for how the task might be solved by the brain.