ELSC Seminar Series

Decision-making unfolds within high-dimensional neural population activity, yet traditional analyses often focus on low-dimensional coding subspaces that represent task-relevant variables. Whether these coding dimensions causally drive decisions or merely reflect them remains unclear. Here, we investigate the mechanistic role of residual dimensions—neural activity patterns orthogonal to coding subspaces—in shaping choice behavior. In the talk I will describe a new class of recurrent neural networks we recently developed that captures both latent coding dynamics and full population activity. Applied to decision-making data, these models reveal that coding dimensions, although accounting for most of the neural variance and correlated with the animal’s choice, have little causal effect on behavior—consistent with recent perturbation experiments. Surprisingly, residual dimensions, while accounting for only a small fraction of the variance, exert a strong causal influence on decisions. We further developed a framework to study the trial-to-trial impact of residual activity on neural dynamics, showing that error trials emerge when fluctuations in residual dimensions steer activity toward incorrect attractors. Extending the model to include thalamocortical loops, we find that inter-area communication operates primarily through residual—not choice-selective—pathways. Together, these results challenge the classical view that decisions are governed solely by low-dimensional coding dynamics, revealing a causal and computational role for residual dimensions in decision-making.