Paper Accepted to ICIP 2022!

Weakly supervised learning approaches can be broadly characterized by three kinds of data supervisions: (1) Incomplete supervision, in which the amount of labeled data is insufficient to train a good model (reduced training scenario); (2) Inexact supervision addresses the lack of information and accuracy about the provided data, and (3) Inaccurate supervision, where the given labels are not always correct as labeling errors (noisy labels) may be present (Zhou, 2017).

Some tasks, such as learning manifolds (aka metric learning), are extremely challenging in weakly supervised scenarios. Usually, such approaches consider the samples' neighborhoods (some distance function is applied) to either map them onto new embeddings or understand their structure. Deep Metric Learning (DML) aims to learn such embeddings using deep learning architectures, with many promising works, such as Nearest Neighbour Gaussian Kernel (NNGK) (Meyer et al., 2018), SoftTriple (Qian et al., 2019), ProxyAnchor (Kim et al., 2020), and Supervised Contrastive (SunCon) (Carlucci et al., 2020), to cite a few.

However, DML approaches do not usually address memorization and are prone to adversarial attacks. Mixup (Zhang et al., 2018) is a simple learning principle that tries to overcome the aforementioned drawbacks. Still, to the best of our knowledge, it has never been applied to the context of metric learning for weakly supervised classification. Mixup is considered an input regularization method and, despite being straightforward, provides promising results (Santos and Papa, 2022).

In this work, we propose three new approaches in the context of DML. We are particularly interested in NNGK due to its robustness and simplicity. As such, we introduce variants that take advantage of Mixup to cope with metric learning in incomplete supervision scenarios.

Do you want to know more about metric learning in incomplete supervision scenarios? Check our paper!