CNN-based image classification on long-tailed CIFAR dataset

[Video Presentation]

1. Problem description

In real-world scenarios, large-scale datasets naturally exhibit imbalanced and long-tailed distributions, where a few categories (majority categories) occupy most of the data while most categories (minority categories) are under-represented [1]. With traditional training formula, commonly used CNN as ResNet [2] has poor performance on such long-tailed datasets. This project will explore the effect of novel loss function and deferred re-weighting (DRW) on CNN training with long-tailed CIFAR datasets.

2. Pre-existing work

Previously, there are several re-sampling methods to deal with imbalanced datasets. However, random re-sampling may lead to overfitting issues.

Another possible way is re-weighting, which optimizes the loss function to increase the loss of the minority class. [3] proposed focal loss to drastically reduce the cross-entropy loss for easy examples (majority categories) while keeping the loss of hard examples (minority categories) at the same level as cross-entropy. figure1

Besides, two-stage training is also effective. The first stage is imbalanced data training. Then transfer learning is conducted with the balanced subset [4].

However, experiments show that simply using re-weighting may not substantially increase the accuracy. A further improvement is to combine two-stage training with re-weighting.

The CNN architecture of this project is ResNet34 [2]. The implementation to deal with long-tailed dataset are based on [8].

3. Main approach

4. Results

Our work from scratch

For grading purposes, here is the list of our work from scratch:

Reference:

[1] Zhang, Yongshun, et al. “Bag of tricks for long-tailed visual recognition with deep convolutional neural networks.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 35. No. 4. 2021.

[2] He, Kaiming, et al. “Deep residual learning for image recognition.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.

[3] Lin, Tsung-Yi, et al. “Focal loss for dense object detection.” Proceedings of the IEEE international conference on computer vision. 2017.

[4] Cui, Yin, et al. “Large scale fine-grained categorization and domain-specific transfer learning.” Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.

[5] https://www.cs.toronto.edu/~kriz/cifar.html

[6] Suykens, Johan AK, and Joos Vandewalle. “Least squares support vector machine classifiers.” Neural processing letters 9.3 (1999): 293-300.

[7] Cao, Kaidi, et al. “Learning imbalanced datasets with label-distribution-aware margin loss.” Advances in neural information processing systems 32 (2019).

[8] https://github.com/kaidic/LDAM-DRW

  
   Back to top