Anastasia Simonoff: Cortex in the Loop: Using Digital Twins to Model Hierarchical Processing in Mouse Visual Cortex

BCCN Berlin / Technische Universität Berlin

Abstract

Visual processing in the brain involves neurons from a hierarchy of visual areas which extract relevant features from the world around us. To best understand how this process occurs, it is essential to identify which features visual cortical neurons extract, i.e., the optimal stimuli which drive neuronal activity. However, this is challenging due to the high dimensionality of visual stimuli. Here, we used recent advances in deep learning to train a digital twin model to reproduce neuronal responses in the mouse visual cortex to visual stimuli. The model was trained on neural and behavioral activity from free-running mice viewing natural images with neurons recorded in primary visual cortex (V1) and a higher visual area, posteromedial cortex (PM), and including anatomical labeling information which neurons were feedforward (V1 to PM) and feedback (PM to V1) projecting neurons. We subsequently used the trained model to predict the most exciting images (MEIs) for each neuron: this is the input that would maximally drive neuronal responses and thus indicates what each neuron prefers in the visual scene. We validated model predictions with independent metrics (sparse noise receptive field mapping) and report factors that contribute to model fitting stability. We used dimensionality reduction to cluster the resulting MEIs and found that the MEIs of V1 and PM cells were more similar within than across areas. We also observed that feedforward and feedback projecting neurons demonstrate MEIs more similar to those of the areas in which they are located in opposed to their initial targets. Additionally, we conducted a comparison of normal and dark-reared mice, revealing smaller receptive fields and less efficient visual encoding in dark-reared animals. Overall, we demonstrate how digital twin models capture more complex, non-linear neuronal preferences, which classical approaches are unable to do. These models provide a powerful tool for investigating feature selectivity across different neuronal populations and offering insights into the hierarchical organization of visual processing in the mammalian brain.

 

 

Additional information:

Master thesis defense

 

Organized by:

Dr. Leopoldo Petreanu, PhD & Prof. Dr. Kerstin Ritter