Projects

Description: This project focuses on developing trustworthy AI methods for low-resource languages, with applications in Automatic Speech Recognition (ASR) and Sign Language Production (SLP). It investigates confidence estimation and uncertainty quantification to assess the reliability of predicted text and sign-language outputs under data-scarce conditions. The project will also develop a human-in-the-loop speech-to-text-to-sign-language platform, where low-confidence predictions can be identified and corrected using language and sign-language resources. The overall goal is to improve the reliability, accessibility, and scalability of AI technologies for Indian languages and Indian Sign Language.

Description: Deep learning has transformed computer vision applications such as object detection and image classification. However, neural networks can be overconfident in their predictions, particularly under changing or unseen real-world conditions. This project focuses on improving the reliability of vision-based machine learning models for autonomous vehicles by developing efficient uncertainty quantification techniques. In particular, the project investigates how to disentangle and quantify aleatoric and epistemic uncertainty using deep feature representations, enabling autonomous systems to identify potentially unreliable predictions and make safer decisions.

Description: The widespread adoption of AI in critical domains such as healthcare, disaster response, natural hazard prediction, and autonomous systems requires reliable mechanisms to assess the trustworthiness of AI predictions. This project focuses on developing an uncertainty quantification framework for multimodal AI models to estimate prediction confidence and improve model calibration. The project investigates confidence estimation, uncertainty disentanglement, and calibration, and aims to develop an uncertainty quantification framework that can be integrated into downstream multimodal AI applications.

Description: The objective of this project is to develop a spoken-language-to-sign-language video synthesis system that converts spoken communication into natural and coherent sign-language videos. The project will focus on temporally smooth sign synthesis and signer identity transformation. The complete pipeline will be optimized to enable seamless spoken-to-sign-language communication.

Description: The objective of this project is to develop an AI-based semantic document similarity platform that can identify and compare semantically related technical documents beyond lexical or keyword-level matching. The project focuses on unsupervised representation learning and dense retrieval to generate meaningful document embeddings and similarity scores, with supporting mechanisms for candidate retrieval, reranking, explainable similarity assessment, and document summarization.

Description: A major challenge that has kept vehicular ad hoc network -VANET still as a subject of research is the security and privacy concerns that it poses in a real-life implementation. One problem is the non-availability of a proper simulation framework for academicians and researchers to investigate VANET and test their algorithms. For this problem, our project built a holistic simulation framework for testing the security of 5G-VANET. The other problem is to build AI based attack detection solution with imbalanced data. For this, our project built AI models to not only detect an attack but also generate a potential attack scenario using generative AI.