Faster Meta Update Strategy for Noise-Robust Deep Learning
About
It has been shown that deep neural networks are prone to overfitting on biased training data. Towards addressing this issue, meta-learning employs a meta model for correcting the training bias. Despite the promising performances, super slow training is currently the bottleneck in the meta learning approaches. In this paper, we introduce a novel Faster Meta Update Strategy (FaMUS) to replace the most expensive step in the meta gradient computation with a faster layer-wise approximation. We empirically find that FaMUS yields not only a reasonably accurate but also a low-variance approximation of the meta gradient. We conduct extensive experiments to verify the proposed method on two tasks. We show our method is able to save two-thirds of the training time while still maintaining the comparable or achieving even better generalization performance. In particular, our method achieves the state-of-the-art performance on both synthetic and realistic noisy labels, and obtains promising performance on long-tailed recognition on standard benchmarks.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Image Classification | Clothing1M (test) | Accuracy74.43 | 546 | |
| Image Classification | ImageNet ILSVRC-2012 (val) | Top-1 Accuracy77 | 405 | |
| Image Classification | ImageNet (val) | Top-1 Accuracy77 | 354 | |
| Image Classification | CIFAR-10 long-tailed (test) | -- | 201 | |
| Image Classification | ILSVRC 2012 (test) | Top-1 Acc77 | 117 | |
| Image Classification | CIFAR-100 (test) | -- | 72 | |
| Image Classification | Webvision (test) | Acc79.4 | 57 | |
| Image Classification | Red Mini-ImageNet (test) | Accuracy51.42 | 51 | |
| Image Classification | CIFAR100-LT (test) | Top-1 Acc (IR=100)46.03 | 45 | |
| Image Classification | WebVision (val) | Top-1 Acc79.4 | 40 |