Search Papers | Poster Sessions | All Posters
Poster C44 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Impact of Stimulus Statistics on Activity Patterns in a Model of Mouse Visual Cortex
Parsa Torabian1, Yinghan Chen1, Shahab Bakhtiari2, Bryan Tripp1 (); 1University of Waterloo, 2Univeristy of Montreal
A number of studies have found that deep networks can be significantly predictive of brain activity in mice, monkeys, and humans. In this work, we used an architecture that closely resembles the mouse visual cortex (MouseNet) and tested the impact of training data statistics on representational similarity with data recorded from the mouse visual system. We used the Unity engine to create eight video datasets with stimulus properties that were either realistic or artificial in three dimensions: environment, motion statistics, and optics of the modelled eye. We used each of these datasets to train the MouseNet model using a self-supervised objective. We found significant, area-dependent variations in similarity scores across different training data conditions. Notably, models trained with realistic environments consistently yielded the highest increases in similarity scores, particularly in higher areas of MouseNet. In contrast, the realistic motion conditions caused area-selective improvements or degradations in similarity scores. Furthermore, across conditions and network instances, we found that self-supervised loss and top-1 accuracy were poorly correlated with similarity to cortical representations. These results are an important step in developing models that more fully account for stimulus-driven mouse brain activity.
Keywords: Dataset Mouse Cortex Representations