Optimizing deep video representation to match brain activity