Global Hardest Example Mining with Prototype-based Triplet Loss

Shan Yang

Abstract: Hard examples are the performance bottleneck of machine learning models, and therefore efficient identification and correct classification of them can significantly improve the model performance. However, most hard example mining schemes search for hard examples in randomly selected mini-batches in each epoch, which often result in local hardest examples and thus sub-optimal performances. Besides, the triplet loss is commonly adopted to explore the mined hard examples by pulling the hard positives close to and pushing the negatives away from the anchor. However, when the anchor in a triplet is an outlier at or close to the cluster boundary, the positive example will be pulled away from the centroid of the cluster, which would result in an incompact cluster, thus inferior performance. To address above challenges, we propose a global hardest example mining with prototype-based triplet loss, which is composed of two major components, namely a Prototype-based Global Hardest Example Miner (GHEM) and a Prototype-based Triplet Loss (pTriplet). First, a global hardest example miner (GHEM) is present to mine firstly the hardest classes on the prototype-based nearest neighbor graph of classes, and then the hardest examples by searching for examples at the cluster boundaries. Second, a prototype-based triplet loss (pTriplet) is developed, which replaces the anchor with an anchor-fused prototype to alleviate the influence of the outlier anchor and provides a normal anchor for triplet loss. Extensive experiments on typical Computer Vision (CV) and Natural Language Processing (NLP) tasks, namely person re-identification and few-shot relation extraction, demonstrated the effectiveness and generalizability of the proposed scheme, which consistently outperforms the-state-of-the-art models. We will publish all source codes of this work on Github for further research explorations.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: General Machine Learning (ie none of the above)