Executive Summary : | The project aims to develop a deep learning model that can recognize classes with very few labeled training examples (few-shot learning) and is robust to label corruption. This is because limited labeled data for few-shot classes may suffer from incorrect labeling, which can lead to poor performance in FSL methods. To address this issue, the project aims to develop a semi-supervised few-shot multi-class recognition model that is robust to label corruption. The problem of this project is to develop a model that can recognize a class using very few training examples while performing joint semi-supervised learning over labeled and unlabeled data, even in the presence of incorrectly labeled training examples. Existing semi-supervised FSL methods often suffer from significant performance degradation due to label corruption, and existing methods do not consider this problem. To address these gaps and issues in the literature, the project will use publicly available benchmark datasets for few-shot recognition for experiments. The model will be built in multiple stages, starting with designing a novel model for FSL with a few-shot learning block (FSLB) for training the network on few-shot data and a label corruption handling block (LCHB) to reduce the effect of label corruption on FSL performance. In stage 2, the model will be extended to incorporate a semi-supervised learning block (SSLB) to enable the FSL model to perform semi-supervised learning on both labeled and unlabeled data while handling FSL using FSLB and label corruption using LCHB. Experiments will be conducted with different proportions of label corruption to evaluate the effectiveness of the proposed approach in dealing with label corruption in this setting. |