Prototypical Classifier for Robust Class-Imbalanced Learning

Abstract

Deep neural networks have been shown to be very powerful methods for many supervised learning tasks. However, they can also easily overfit to training set biases, i.e., label noise and class imbalance. While both learning with noisy labels and class-imbalanced learning have received tremendous attention, existing works mainly focus on one of these two training set biases. To fill the gap, we propose Prototypical Classifier, which does not require fitting additional parameters given the embedding network. Unlike conventional classifiers that are biased towards head classes, Prototypical Classifier produces balanced and comparable predictions for all classes even though the training set is class-imbalanced. By leveraging this appealing property, we can easily detect noisy labels by thresholding the confidence scores predicted by Prototypical Classifier, where the threshold is dynamically adjusted through the iteration. A sample reweighting strategy is then applied to mitigate the influence of noisy labels. We test our method on both benchmark and real-world datasets, observing that Prototypical Classifier obtains substaintial improvements compared with state of the arts.

T. Wei and J.-X. Shi—Co-first authors. This work was done when Tong Wei was a student at Nanjing University.

Appendices

A Ablations on Dynamic Threshold

Figure 6 shows a comparison of fixed threshold and the dynamic threshold $\tau _t$ with $\tau _0 = 0.1$. We consider both exponential scheduler controlled by $\gamma$ and linear scheduler controlled by the threshold of last iteration $\tau _T$.

We test the performance of different choice of parameters and the results are reported in Table 6. From the results, we have two observations: i) when using fixed threshold or the dynamic threshold grows too slow, performance drops in the last iterations because many noisy labels are incorrectly flagged as clean; and ii) when dynamic threshold grows too fast, the network cannot achieve best performance, because many clean labels are incorrectly flagged as noisy.

Fig. 6.

Comparison of fixed threshold and dynamic threshold. Fix threshold $\tau =0.1$, exponential dynamic threshold $\tau _t=0.1\gamma ^{t}$ and linear dynamic threshold $\tau _t = 0.1 + \frac{\tau _T-0.1 }{T}t$.

Table 6. Test accuracy (%) on CIFAR-10-LT with imbalance factor 100 and noise ratio 50%.

Table 7. Test accuracy (%) on clean datasets with different imbalanced factor.

B Results on Clean Datasets

Although our method is particularly designed learning with noisy labels, it is interesting to study its performance on clean but class-imbalanced datasets. In this experiment, we do not use sample re-weighting and label noise correction. We report the results in Table 7. For fair comparison, we do not apply AugMix in this experiment. Intriguingly, Prototypical Classifier consistently outperforms all baselines by a large margin, showing the superiority of our proposed representation learning method.