In the first part of the talk, I will propose sparsemax, a new activation function similar to the traditional softmax, but able to output sparse probabilities. After deriving its properties, I will show how its Jacobian can be efficiently computed, enabling its use in a network trained with backpropagation. Then, I will propose a new smooth and convex loss function which is the sparsemax analogue of the logistic loss, revealing an unexpected connection with the Huber classification loss. I will show promising empirical results in multi-label classification problems and in attention-based neural networks for natural language inference.

In the second part, I will introduce constrained softmax, another activation function that allows imposing upper bound constraints on attention probabilities. Based on this activation, I will introduce a novel neural end-to-end differentiable easy-first decoder that learns to solve sequence tagging tasks in a flexible order. The decoder iteratively updates a "sketch" of the predictions over the sequence. The proposed models compare favourably to BILSTM taggers on three sequence tagging tasks.

This is joint work with Ramon Astudillo and Julia Kreutzer.