Connectivity predicts where computation forms and where it can be changed

In recurrent networks trained on 15 cognitive tasks, connectivity-defined unit classes predict dynamical motifs and intervention effects, including a 19.1× larger target shift from modulating the highest-leverage class.

• Why it matters: Circuit connectivity is known to constrain population dynamics, but it has been unclear how mesoscale structure in recurrent weights determines which computations emerge and how they can be altered.
• What they did: The study groups units by the aggregate input each unit receives from other groups, then uses recurrent weights alone to identify connectivity-defined classes and compare their contributions to dynamical motifs across tasks.
• The result: Units in the same class contribute similarly to the same dynamical motifs, class interactions recover motif geometry without fitted parameters, and state-dependent gain reconfigures effective dynamics; modulating the class with greatest dynamical leverage shifts a target component 19.1 times farther than modulating a typical class.

AI summary of the abstract · may contain errors · not medical advice · read the paper.