Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Unsupervised Identification of Behaviorally-relevant States in Biological and Artificial Neural Systems

Arman Behrad¹, Mohammad Taha Fakharian², Christian Beste, Shervin Safavi¹; ¹Technische Universität Dresden, ²University of Tehran, University of Tehran

Presenter: Arman Behrad

Identifying distinct neural dynamics corresponding to cognitive states and their transitions is crucial for understanding the neural machinery of cognitive functions in both biological and artificial intelligent systems. However, conventional methods for state identification constrain the analysis by relying on predefined state labels. Here, we introduce a novel unsupervised approach to robustly detect behaviorally-relevant state transitions without prior assumptions or knowledge about behavioral labels. We assume that each state has a characteristic dynamics within each state (with minimal variation), but triggered by behavioral demands, transitions to other states (with different characteristic dynamics). Therefore, comparing neural dynamics across time, should provide us key information about state transitions. Based on this idea, we developed Moving Window Dynamical Similarity Analysis (MoDSA) for an unbiased detection of state transitions in neural systems. We validated our method on biological neural data recorded from macaque area V4 during selective attention tasks, and data from diverse recurrent neural networks trained on context-dependent decision-making tasks. We demonstrate that our method can identify behaviorally meaningful states purely based on neural dynamics, in both domains of artificial and biological neural systems.

Topic Area: Brain Networks & Neural Dynamics

Extended Abstract: Full Text PDF