Seminar by Caroline Haimerl

Friday 12th September 2025, at 11:00 AM, at the INT, Henri Gastaut meeting room

Caroline Haimerl (Champalimaud Center) invited by Fanny Cazettes

Modeling the multiarea computations underlying multiscale behavior

Abstract: How flexible adaptive behavior arises from slowly learned cortical representations and fast task learning remains a central challenge in neuroscience. Consider goal-directed navigation: we quickly adjust trajectories to avoid accidents (ms-sec), rethink routes given blocked paths (min-days), and adapt habits – like biking – over seasons. While seemingly effortless, this adaptability involves vast numbers of decisions at multiple spatiotemporal scales, coordinated across distributed brain circuits. These complex interactions make it difficult to infer from the necessarily partial observations of neural activity, how local representations in individual brain areas contribute to coherent behavior across scales. My theoretical frameworks and computational models aim to link hypotheses about joint objectives and intermediate computations with multi-area neural and behavioral data. I will present past and ongoing work modeling hierarchical cortical representation learning and parallel task-structure learning to explain the multiscale organization of behavior, and integrating biologically inspired learning mechanisms into hierarchical neural network architectures to study continual task adaptation. Leveraging these models to guide the analyses of neural population activity, we showed that (1) millisecond fluctuations in gain modulation of hierarchical cortical representations could underlie fast task adaptation, and (2) integrating parallel reinforcement learning algorithms within hierarchical state-action representations captures both multiscale task learning and neural coding in striatum. With these modeling frameworks and large-scale data analysis, I aim to link normative principles with mechanistic hypotheses to advance our understanding of adaptive behavior.