Training neural networks using a clustering loss

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network. One of the methods includes obtaining a batch of training items and a ground truth assignment; processing the training items in the batch using the neural network to generate respective embeddings for each of the training items; and adjusting the current values of the network parameters by performing an iteration of a neural network training procedure to optimize an objective function that penalizes the neural network for generating embeddings that do not result in, for each possible clustering assignment other than the ground truth assignment, the oracle clustering score being higher than a clustering score for the possible clustering assignment by at least a structured margin between the possible clustering assignment and the ground truth assignment.