Neuroscientist and AI expert Andrew Saxe explains the importance of constraints and inference in neural networks, specifically the need for coordinated transformations and stable connections between adjacent layers.