> suggests that gradient descent optimization doesn't cleverly pick good internal representations of the data. How so? The existence of equivalent kernels does

> suggests that gradient descent optimization doesn't cleverly pick good internal representations of the data. How so? The existence of equivalent kernels doesn't imply they aren't clever or good.