Distributional Robustness Loss for Long-tail Learning

Samuel, Dvir; Chechik, Gal

doi:10.48550/arxiv.2104.03066

Cited by 5 publications

(12 citation statements)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We further demonstrate our model's ability to integrate new, previously unseen classes with strong recognition accuracy, and provide an analysis of our model training strategy so as to facilitate knowledge transfer. Nonetheless, our approach is not tied to a specific backbone training approach, and further benefits could be obtained by combining our strategy with backbone training methods comprising auxiliary losses that optimise for class separability [7,26].…”

Section: Discussionmentioning

confidence: 99%

“…It has been shown that backbone capacity and additional training losses can substantially improve backbone quality and consequently classification performance [7,15,26]. As our approach is independent from the backbone training process, we train our model using the setting that matches most pre-existing approaches.…”

Section: Comparison To State Of the Art Methodsmentioning

confidence: 99%

“…Zhong et al [35] rely on label smoothing and mixup strategies, while Cui et al [7] combine cross-entropy with supervised contrastive learning, achieving stronger representations at the cost of more expensive training. Finally, closest to our work, Samuel et al [26] propose to rely on a prototype-based auxiliary loss, where class prototypes are computed at the beginning of every epoch during training. Introducing learnable class specific uncertainty parameters and weights yields stronger backbone representations, yet classification still relies on learned classifier weights.…”

Section: Related Workmentioning

confidence: 99%

“…Our work focuses on this issue, and is not tied to a rebalancing strategy. Therefore, it provides complementary benefits to approaches that focus on backbone training improvements [7,24,26], as discussed above.…”

Section: Related Workmentioning

confidence: 99%

See 3 more Smart Citations

Long-tail Recognition via Compositional Knowledge Transfer

Parisot¹,

Esperança²,

McDonagh³

et al. 2021

Preprint

View full text Add to dashboard Cite

In this work, we introduce a novel strategy for long-tail recognition that addresses the tail classes' few-shot problem via training-free knowledge transfer. Our objective is to transfer knowledge acquired from information-rich common classes to semantically similar, and yet data-hungry, rare classes in order to obtain stronger tail class representations. We leverage the fact that class prototypes and learned cosine classifiers provide two different, complementary representations of class cluster centres in feature space, and use an attention mechanism to select and recompose learned classifier features from common classes to obtain higher quality rare class representations. Our knowledge transfer process is training free, reducing overfitting risks, and can afford continual extension of classifiers to new classes. Experiments show that our approach can achieve significant performance boosts on rare classes while maintaining robust common class performance, outperforming directly comparable state-of-the-art models.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Comparison To State Of the Art Methodsmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

See 2 more Smart Citations

Long-tail Recognition via Compositional Knowledge Transfer

Parisot¹,

Esperança²,

McDonagh³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…After that, some decoupling methods [21,54] also can be regarded as transferring head classes frozen feature to tail classes when fine-tuning classifiers. Recently, some studies [7,20,39,46,52] also transfer the representation learned by contrastive learning or selfsupervised learning for long-tailed problems.…”

Section: Long-tailed Visual Recognitionmentioning

confidence: 99%

VL-LTR: Learning Class-wise Visual-Linguistic Representation for Long-Tailed Visual Recognition

Tian¹,

Wang²,

Zhu³

et al. 2021

Preprint

View full text Add to dashboard Cite

Deep learning-based models encounter challenges when processing long-tailed data in the real world. Existing solutions usually employ some balancing strategies or transfer learning to deal with the class imbalance problem, based on the image modality. In this work, we present a visuallinguistic long-tailed recognition framework, termed VL-LTR, and conduct empirical studies on the benefits of introducing text modality for long-tailed recognition (LTR). Compared to existing approaches, the proposed VL-LTR has the following merits. (1) Our method can not only learn visual representation from images but also learn corresponding linguistic representation from noisy classlevel text descriptions collected from the Internet; (2) Our method can effectively use the learned visual-linguistic representation to improve the visual recognition performance, especially for classes with fewer image samples. We also conduct extensive experiments and set the new state-of-theart performance on widely-used LTR benchmarks. Notably, our method achieves 77.2% overall accuracy on ImageNet-LT, which significantly outperforms the previous best method by over 17 points, and is close to the prevailing performance training on the full ImageNet. Code shall be released.

show abstract

Robust Long-Tailed Learning under Label Noise

Wei¹,

Shi²,

Tu³

et al. 2021

Preprint

View full text Add to dashboard Cite

Long-tailed learning has attracted much attention recently, with the goal of improving generalisation for tail classes. Most existing works use supervised learning without considering the prevailing noise in the training dataset. To move long-tailed learning towards more realistic scenarios, this work investigates the label noise problem under long-tailed label distribution. We first observe the negative impact of noisy labels on the performance of existing methods, revealing the intrinsic challenges of this problem. As the most commonly used approach to cope with noisy labels in previous literature, we then find that the small-loss trick fails under long-tailed label distribution. The reason is that deep neural networks cannot distinguish correctly-labeled and mislabeled examples on tail classes. To overcome this limitation, we establish a new prototypical noise detection method by designing a distance-based metric that is resistant to label noise. Based on the above findings, we propose a robust framework, ROLT, that realizes noise detection for long-tailed learning, followed by soft pseudo-labeling via both label smoothing and diverse label guessing. Moreover, our framework can naturally leverage semi-supervised learning algorithms to further improve the generalisation. Extensive experiments on benchmark and real-world datasets demonstrate the superiority of our methods over existing baselines. In particular, our method outperforms DivideMix by 3% in test accuracy. Source code will be released soon.

show abstract

Distributional Robustness Loss for Long-tail Learning

Cited by 5 publications

References 35 publications

Long-tail Recognition via Compositional Knowledge Transfer

Long-tail Recognition via Compositional Knowledge Transfer

VL-LTR: Learning Class-wise Visual-Linguistic Representation for Long-Tailed Visual Recognition

Robust Long-Tailed Learning under Label Noise

Contact Info

Product

Resources

About