A Computation-through-Dynamics Benchmark extended to Neural ODE Models of Perceptual Decision-Making

Abstract

Verifying the dynamical similarity between data-drained deep learning models and the biological circuits they aim to replicate remains a significant challenge. Benchmarking methods that evaluate such models based on their underlying dynamical systems, rather than only their output performance, are thus highly desirable. In this work, we validate and extend one such recently proposed method: the Computation through Dynamics Benchmark (CtD-B). In the context of models that solve the Poisson-clicks task (a perceptual decision-making cognitive task), we test existing CtD-B metrics and find that functional similarity measures — Rate R2 and Dynamical Systems Alignment (DSA/co-BPS) — are robust across models, but representational metrics — State R2 and Cycle-Convergence (Cycle-Con) — are reliable for low-dimensional models. Leveraging dynamical systems theory, we extend the analysis function of the benchmark to consider local similarity: fixed points and timescales in both task-trained (TT) and data-driven (DD) models. Notably, we find that DD models can fit observed data without preserving the characteristic timescales of TT solutions.