Faster Gaze Prediction: Dense Networks and Fisher Pruning

This lecture discusses the development of more efficient architectures for saliency prediction by leveraging deep representations pre-trained for object recognition. The instructor presents a greedy pruning method called Fisher pruning, combined with knowledge distillation, to achieve faster single-image gaze prediction. The lecture emphasizes the importance of speeding up gaze prediction for real-world applications and video saliency models. Additionally, the lecture explores the challenges and recent developments in statistical learning, bias, ground truth, and ethics in critical data studies.