CAIM talk by Guillaume Bellec: Perturbation testing to validate deep learning models of cortical computation
Date: February 25th, 2026, 1pm
Location: Medical University of Vienna, Anna Spiegel Research Building, Seminar Room Level 3
Speaker: Guillaume Bellec
Title: Perturbation testing to validate deep learning models of cortical computation
Abstract: Deep learning models can accurately fit neural activity recordings, but do they capture the causal mechanisms of cortical computation? I argue that perturbation testing (e.g. predicting the behavioral effects of targeted neural stimulation) provides a decisive validation metric and methodology. I define three levels of perturbation testing with increasing experimental complexity. At Level 1, I review recent works (ours included) combining deep learning with interpretable stimulation models that capture biophysical cortical structure (cell-type connectivity, feedback loops, topographic layouts). When these models are decisive to pass out-of-distribution perturbation tests, it provides validated insights about the functional role of cortical mesoscopic structures. Looking further, Levels 2 and 3 anticipate deep generative models will be trained on dense perturbation datasets and tested on closed-loop experiments. I define evaluation metrics and open conceptual problems towards reverse engineering cortical computation from these emerging digital twins.
This is part of the CAIM Talks series.