Example the bias of CNNs to exploit high-frequency components in image might be good for CV benchmark performance, if you feed the same representations to a robotic policy then it is clearly sub-optimal, shape is more important here ....

Comments