MTech (Rresearch) Colloquium

Name of the Student:      Ahmad Arfeen

Guide:                             Prof. Soma Biswas

Date And Day:                21st April, 2022, Thursday

Time:                               11:00 am

Venue:                              EE, MMCR

Title : Data Efficient Domain Generalization

Abstract: For the task of image classification, in general, the test data is assumed to come from the same distribution as the training data. But this may not always hold in real-life scenarios. For example, in night-time surveillance, we may need to classify images captured using NIR cameras, but the available model has been trained on RGB images. Domain generalization (DG) addresses the problem of generalizing classification performance across any unknown domain, by leveraging training samples from multiple source domains. In this thesis, we address two challenging scenarios for the DG task. Currently, the training process of majority of the state-of-the-art DG-methods is dependent on a large amount of labeled data. This restricts the application of the models in many real-world scenarios, where collecting and annotating a large dataset is an expensive and difficult task. 

Thus, as the first contribution, we address the problem of Semi-supervised Domain Generalization (SSDG), where the training set contains only a few labeled data, in addition to a large number of unlabeled data from multiple domains. This is relatively unexplored in literature and poses a considerable challenge to the state-of-the-art DG models, since their performance degrades under such condition. To address this scenario, we propose a novel Selective Mixing and Voting Network (SMV-Net), which effectively extracts useful knowledge from the set of unlabeled training data, available to the model. Specifically, we propose a mixing strategy on selected unlabeled samples on which the model is confident about their predicted class labels to achieve a domain-invariant representation of the data, which generalizes effectively across any unseen domain. Extensive experiments on two popular DG-datasets demonstrate the usefulness of the proposed framework. 

The second contribution of this thesis is a novel approach for the task of Zero-Shot Domain Generalization (ZSDG). This is very challenging since the query data can belong to an unseen class as well as unseen domain. For this task, we address the challenge of class imbalance by learning class-specific classifier margins, which not only maintain the semantic relationship of the classes in the embedding space, but is also discriminative, and thus improves the classification performance on the test data. Extensive experiments on multiple datasets justify the effectiveness of the proposed approach.

ALL ARE WELCOME