Expeditious Saliency-guided Mix-up through Random Gradient Thresholding

9 Dec 2022  ·  Minh-Long Luu, Zeyi Huang, Eric P. Xing, Yong Jae Lee, Haohan Wang ·

Mix-up training approaches have proven to be effective in improving the generalization ability of Deep Neural Networks. Over the years, the research community expands mix-up methods into two directions, with extensive efforts to improve saliency-guided procedures but minimal focus on the arbitrary path, leaving the randomization domain unexplored. In this paper, inspired by the superior qualities of each direction over one another, we introduce a novel method that lies at the junction of the two routes. By combining the best elements of randomness and saliency utilization, our method balances speed, simplicity, and accuracy. We name our method R-Mix following the concept of "Random Mix-up". We demonstrate its effectiveness in generalization, weakly supervised object localization, calibration, and robustness to adversarial attacks. Finally, in order to address the question of whether there exists a better decision protocol, we train a Reinforcement Learning agent that decides the mix-up policies based on the classifier's performance, reducing dependency on human-designed objectives and hyperparameter tuning. Extensive experiments further show that the agent is capable of performing at the cutting-edge level, laying the foundation for a fully automatic mix-up. Our code is released at [https://github.com/minhlong94/Random-Mixup].

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Classifier calibration CIFAR-100 R-Mix (PreActResNet-18) Expected Calibration Error 3.73 # 1
Image Classification CIFAR-100 RL-Mix (WideResNet 28-10) Percentage correct 84.9 # 73
Image Classification CIFAR-100 R-Mix (WideResNet 28-10) Percentage correct 85 # 70
Image Classification CIFAR-100 PreActResNet-18 + CutMix (OneCycleLR scheduler) Percentage correct 80.6 # 122
Image Classification CIFAR-100 RL-Mix (PreActResNet-18) Percentage correct 80.75 # 121
Image Classification CIFAR-100 R-Mix (PreActResNet-18) Percentage correct 81.49 # 117
Image Classification CIFAR-100 WideResNet 16-8 + CutMix (OneCycleLR scheduler) Percentage correct 81.79 # 111
Image Classification CIFAR-100 RL-Mix (WideResNet 16-8) Percentage correct 82.16 # 107
Image Classification CIFAR-100 ResNeXt 29-4-24 + CutMix (OneCycleLR scheduler) Percentage correct 82.3 # 105
Image Classification CIFAR-100 R-Mix (WideResNet 16-8) Percentage correct 82.32 # 104
Image Classification CIFAR-100 RL-Mix (ResNeXt 29-4-24) Percentage correct 82.43 # 102
Image Classification CIFAR-100 R-Mix (ResNeXt 29-4-24) Percentage correct 83.02 # 93
Image Classification CIFAR-100 WideResNet 28-10 + CutMix (OneCycleLR scheduler) Percentage correct 83.97 # 81
Image Classification ImageNet R-Mix (ResNet-50) Top 1 Accuracy 77.39% # 812
Weakly-Supervised Object Localization ImageNet R-Mix (ResNet-50) Top-1 Localization Accuracy 55.58 # 3

Methods