| Abstract |
I will present a differentiable sorting mechanism inspired by the quicksort algorithm. Where previous approaches to learnable sorting designed inductive biases for permutations, we take a more direct approach. We implement a mechanism which sorts a given sequence of tensors by a matching sequence of scalar key values. The sorting operation itself does not depend on learnable parameters, but it does allow a gradient to propagate back to the keys. This allows us to effectively train a network to compute the keys based only on examples of sorted data. |