Benchmarking Geometric Assumptions in Active Learning for Non-Normal Neural Dynamics

As systems neuroscience shifts toward active, closed-loop control, efficient system identification within limited experimental sessions is critical. However, emerging active learning strategies lack a standardized framework to compare their differing geometric assumptions and objectives. To address this, we introduce a common formalism for A-optimal active learning (AL) of linear neural dynamics, alongside a comprehensive benchmark designed to evaluate active learning approaches for neural population dynamics. We systematically evaluate state-of-the-art sampling methodologies under challenging, biologically relevant conditions: low intrinsic dimensionality and non-normal dynamics. Informed by this evaluation, we propose a novel hybrid active learning approach that achieves significantly faster convergence to latent dynamics, particularly in the non-normal case, compared to existing approaches. Ultimately, this benchmark provides critical insights into the assumptions and limitations of current active learning approaches, synthesizing them into a robust approach suitable for identifying computations performed by biological neural networks.