Cleaning label noise with vision-language models

Chen Feng, Ioannis Patras

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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.

Keywords: Label noise, Sample selection

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose an effective dataset cleaning approach utilizing the vision-language model CLIP, which mitigates the biases of the visual modality and alleviates self-confirmatory bias.

Abstract:

Current mainstream methods for learning with noisy labels often rely on sample selection, such as the common 'small-loss' strategy that considers samples with smaller losses as clean. Following this, most research focuses on developing more robust sample selection strategies. However, they are still influenced by problems such as the 'self-confirmation bias', which stems from their reliance on the in-training model. Furthermore, relying solely on visual information for sample selection can introduce biases and challenges, such as the common issue of 'hard noise', where samples are erroneously labeled as semantically similar categories.
To address these challenges, this paper proposes using the popular vision-language model CLIP for sample selection. Leveraging CLIP, a pre-trained model, can effectively mitigate self-confirmation bias. Additionally, CLIP's distinctive language modality supplements potential biases introduced by relying solely on visual information for sample selection.
Specifically, we introduce the \textit{CLIPSelector}, which utilizes both the CLIP's zero-shot classifier and an easily-inducible classifier based on its vision encoder and noisy labels for sample selection. We theoretically and empirically demonstrate the unique advantages of the \textit{CLIPSelector}.
To evaluate its effectiveness on existing benchmarks, we further introduce a semi-supervised learning method called \textit{MixFix}, tailored for noisy datasets. \textit{MixFix} leverages the subset selected by the \textit{CLIPSelector} and gradually introduces missing clean samples and re-labeled noisy samples based on different thresholds.
In comparison to current hybrid methods involving iterative sample selection and multiple off-the-shelf techniques like model co-training, our approach simplifies the process.
Nonetheless, our approach achieves competitive or superior performance across various benchmarks, including datasets with synthetic and real-world noise.
Code will be released upon acceptance.

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.

Submission Number: 3300