Recurrent network models of motor cortex

A productive way to test the dynamical-systems view is to train a recurrent neural network to produce the muscle activity or kinematics of a task and then compare its internal workings with cortex. Such task-trained networks, given only inputs and desired outputs, spontaneously develop many of the population features seen in data — rotational structure, a large condition-independent signal at movement onset, and heterogeneous, multiphasic single-unit responses — supporting the idea that these are generic consequences of a network solving the movement-generation problem rather than idiosyncrasies of biology. The trained network then serves as a hypothesis whose mechanism can be dissected by locating and linearizing its fixed points.

These models constrain theories but do not settle them. Many different networks, and many different training regularizations, can reproduce the same neural signatures, so a good match is evidence for a class of solutions rather than proof that cortex uses a particular one; the choices of architecture, noise, and cost function strongly shape what emerges. Used carefully — as a source of falsifiable predictions and a tool for reverse-engineering computation — recurrent models are among the most useful bridges between recorded population dynamics and mechanism, provided their degeneracy is acknowledged.