The shared geometry of biological and recurrent neural network dynamics

Recent studies have proposed that the activity of both recurrent and biological neural networks is constrained to a task manifold. This manifold is often intrinsically lower dimensional than the full neural state space, and its geometry is directly related to task computations. In parallel, there has been a surge of interest in training RNNs to solve behavioural tasks in order to generate testable hypotheses regarding the computations employed by neural circuits. However, means of using task-trained RNNs to test hypotheses regarding the geometry of neural data manifolds are currently lacking.

In the present work, we develop a method to address this need. Our method consists of aligning the RNN and neural data manifolds, while accounting for trial-to-trial variability in trajectories on the data manifold by controlling and time-warping the trajectories of the RNN. We apply our method to two large-scale datasets. First, using recordings from prefrontal cortex during a contextual decision making task, we show that our method uncovers task-relevant structure in the neural data that is otherwise hidden in the high-dimensional data. Second, using motor and premotor cortical activity in a delayed center-out reach task, we demonstrate that by varying the architecture of the RNN and the design of the task it is trained on, distinct hypotheses can be generated regarding the origin of motor preparatory activity and rotational dynamics that can be tested in the neural data. Together, our results illustrate how hypotheses regarding the latent computations employed by neural circuits to solve behavioural tasks can be framed and tested by comparing the geometry of recurrent and biological neural network activity.