Conference on Cognitive Computational Neuroscience

How can we tell if two neural networks are performing the same computations? This question has grown increasingly important as the capabilities of artificial neural network models and experimental methods for recording neural data have improved. Thus far, most attempts to answer this question have emphasized how neural networks, both artificial and biological, process information in terms of the spatial geometry of their neural activity. However, this approach does not account for how information is transformed over time, which is essential for understanding biological circuits. To address this issue, we developed a data-driven method called Dynamic Mode Representational Similarity Analysis (DMRSA). DMRSA utilizes a high-dimensional embedding to identify spatiotemporally coherent features of two nonlinear dynamical systems. These features, referred to as Koopman Modes, describe core dynamic patterns. Subsequently, a statistical shape analysis is used to compare these modes, thereby assessing the similarity between the systems' dynamics. Our results demonstrate that DMRSA effectively identifies the dynamic structure of neural computations, whereas standard geometric methods fall short. DMRSA therefore has significant relevance for neuroscience, where knowledge of the underlying dynamics can only be inferred through measurements. Consequently, DMRSA opens the door to novel data-driven analyses of the temporal structure of neural computation.