Weakly supervised learning is an essential problem in computer vision tasks, such as image classification, object recognition, etc., because it is expected to work in the scenarios where a large dataset with clean labels is not available.
- Implementing a noise regularizing CNN model and Weakly supervised GAN that can classify images with noisy labels.
- For training data: Split the dataset into clean labels (<10%) and dirty labels (>90%). For the dirty labels, randomly shuffle the labels among images.
- Replicate the noise regularizing model shown in https://openaccess.thecvf.com/content_CVPR_2019/papers/Hu_Weakly_Supervised_Image_Classification_Through_Noise_Regularization_CVPR_2019_paper.pdf and train your dataset.
- Replicate the WSGAN shown in https://arxiv.org/pdf/2111.14605.pdf and train your dataset.
- Compare the performance on your test data and explore the difference between two methods.
- Develop a model which can leverage the adversarial learning architecture in GAN to replace the regularizing network to improve performance. Test your model by change the split level for clean labels starting from 10% and getting down as low as 5%. Find the cut off for which the model performance stops improving.
- Dataset: https://pytorch.org/vision/stable/generated/torchvision.datasets.FashionMNIST.html (or any multilabel dataset of your choice)