Locality Sensitive Batch Selection for Triplet Networks

Triplet networks are deep metric learners which learn to optimise a feature space using similarity knowledge gained from training on triplets of data simultaneously. The architecture relies on the triplet loss function to optimise its weights based upon the distance between triplet members. Composition of input triplets therefore directly impacts the quality of the learned representations, meaning that a training scheme which optimises their formation is crucial. However, an exhaustive search for the best triplets is prohibitive unless the search for triplets is confined to smaller training regions or batches. Accordingly, current triplet mining approaches use informed selection applied only to a random minibatch, but the resulting view fails to exploit areas of complexity in the feature space. In this work, we introduce a locality-sensitive batching strategy, which uses the locality of examples to create batches as an alternative to the commonly adopted randomly minibatching. Our results demonstrate this method to offer better performance on three image and two text classification tasks with statistical significance. Importantly most of these gains are incrementally realised with as little as 25% of the training iterations.