查看文章

Siamese Neural Networks (SNNs) are deep metric learners that use paired instance comparisons to learn similarity. The neural feature maps learnt in this way provide useful representations for classification tasks. Learning in SNNs is not reliant on explicit class knowledge; instead they require knowledge about the relationship between pairs. Though often ignored, we have found that appropriate pair selection is crucial to maximising training efficiency, particularly in scenarios where examples are limited. In this paper, we study the role of informed pair selection and propose a 2-phased strategy of exploration and exploitation. Random sampling provides the needed coverage for exploration, while areas of uncertainty modeled by neighbourhood properties of the pairs drive exploitation. We adopt curriculum learning to organise the ordering of pairs at training time using similarity knowledge as a heuristic for …