Grant: Canada Research Chairs

Computer vision has been revolutionized by the development of convolutional neural network (CNN) models that were inspired by the way the human brain processes visual information. These same CNN models have set the current state-of-the-art at predicting visually evoked neural activity in the retina and the brain’s visual cortex, and recent work from Zylberberg and colleagues demonstrated that training CNN models to mimic the brain’s visually evoked neural activity patterns causes those models to perform better at object recognition tasks. This body of research emphasizes that improved computational models of the visual system drive advances in both visual neuroscience and computer vision. Despite these advances, current computational models of the visual system are limited in their ability to predict neural activity in new (previously unseen) individuals, and they fail to predict visually evoked neural activity under variable lighting conditions such as those that people experience when driving through the city on a sunny day. Dr. Zylberberg’s research program aims to overcome these limitations by applying new approaches from artificial intelligence (AI) known as foundation model approaches, and by incorporating into deep learning models more bio-realistic components that adapt to changing lighting conditions the same way the visual system does. Finally, Dr. Zylberberg will apply the improvements in visual system modeling that result from this research to his ongoing efforts to use neural data to develop better computer vision systems. Altogether, this research program will improve computational visual system models for application in both neuroscience and computer vision.