Scholar
Susmit Agrawal
Google Scholar ID: dwx-5E0AAAAJ
PhD Candidate at IMPRS-IS
Citations & Impact
Citations
198
H-index
6
i10-index
6
Publications
15
Co-authors
31
list available
Publications
Resume (English only)
Academic Achievements
- Insufficient information
Research Experience
- Collaborated with Prof. Chirag Agarwal at the University of Virginia. Master's thesis was based on adapting pre-trained models for downstream tasks - using such models for concept-based explanations, applying them in continual learning settings, and analyzing fine-tuning methods. Briefly worked on understanding and modelling human behavior to create systems that can understand user requirements and adapt their outputs accordingly. Worked with Prof. R. Venkatesh Babu at the Indian Institute of Science on multiple domains in Computer Vision and Computational Photography - particularly High Dynamic Range Imaging and Representation Learning. Briefly worked at Qualcomm R&D on no-reference image quality analysis, image super-resolution and image-to-image translation under hardware constraints. Interned with the Adobe Media Data Science and Research (MDSR) Team, where collaborated with Balaji Krishnamurthy and Yaman Kumar on aligning LLMs to human opinions.
Education
- Currently a PhD candidate at the Bethge Lab at IMPRS-IS. Previously, a 3-year M.Tech. (RA) student at the Indian Institute of Technology - Hyderabad, advised by Prof. Vineeth N. Balasubramanian, working on multimodal learning and explainable neural architectures.
Background
- Interested in building systems that take inspiration from natural intelligence, solving problems such as fixation prediction, continual learning, composition-based learning, or consolidation of multimodal input streams - all of which the human brain solves effortlessly. Also interested in the various aspects of memory - its relation to neural architectures, its ability to link disparate knowledge sources, its emergence in large pre-trained models, and its importance in enabling rapid learning of different tasks or cross-task knowledge transfer.
Miscellany
- Keywords: NeuroAI, Vision, Continual Learning, Concept-based Explainability, Multimodal Learning, Memory, Representation Learning
