In fact, it had redundant internal representations of the goal square! Due to how CNNs work, that should be literally meaningless!
What does this mean? I don’t know as much about CNNs as you—are you saying that their architecture allows for the reuse of internal representations, such that redundancy should never arise? Or are you saying that the goal square shouldn’t be representable by this architecture?
What does this mean? I don’t know as much about CNNs as you—are you saying that their architecture allows for the reuse of internal representations, such that redundancy should never arise? Or are you saying that the goal square shouldn’t be representable by this architecture?