Research Focus
My research focuses on application of neural and symbolic methods to problems in language (textual) modality as well as those at the intersection of language and vision. Specifically, I am interested in exploring the adoption of commonsense understanding and reasoning to vision+language tasks. One could summarize my research focus as:Language + Vision + KnowledgeMaster's Research
During my Masters, I primarily worked on:- Commonsense-driven grounding of natural language queries to raw video data: Researching neuro-symbolic methodologies for language grounding in raw videos (Accepted to AAAI'24!🎉)
- Semantic contrastive pre-training: Multimodal pre-training using multimodal transformers
Academic Service
News
Publications
Experience
Teaching
Research
PLAN (Perception + LANguage) Lab, Virginia Tech
Graduate Research Assistant
Research: Investigating self-supervised methods for language grounding in videos by leveraging background commonsense knowledge.Center for Pattern Recognition and Machine Intelligence, PES University
Undergraduate Research Assistant
Research: Facial Recognition driven by keypoint feature detectors and descriptors; Semantic Segmentation of clothing imagesProfessional Experience
Bloomberg LP, New York
ML Intern
🛠 PyTorch, PythonMorgan Stanley, Bangalore
Technology Associate
🛠 Python, Java, Apache Solr, Lucidworks Fusion, Kafka, KSQLMorgan Stanley, Bangalore
Technology Analyst Intern
🛠 Python (Flask), Angular, DB2Morgan Stanley, Bangalore
Summer Analyst
🛠 Java, Angular, D3.jsMapmyIndia, Bangalore
Machine Learning Intern
🛠 Python, Tensoflow, LinuxSelected Projects
Formulated Visual Question Answering (VQA) as a Visual Entailment (VE) task by using an image as the premise and joint representation of question and answer as a hypothesis.
Performed detailed experiments using both VQA and VE datasets to assess their inherent properties.
Identified and interpreted robustness and latent biases in multimodal datasets and models for VE/VQA.
Designed and developed a description-driven face generation module capable of generating and making corrections to a face through iterative natural language feedback
Additionally developed an end-to-end pipeline a speech-to-text module for ergonomic interfacing
The multi-pass nature enables a dialog-based communication with the generator model
Published at the 32nd IEEE International Conference on Tools with Artificial Intelligence (ICTAI 2020) held in Baltimore, MD (Acceptance rate: 25%)
Proposed a methodology for token-wise estimation of degree of emphasis using BiLSTM-CRF with Label Distribution Learning
Published at the 4th International Conference on Advanced Informatics for Computing Research (ICAICR 2020)
Designed a methodology for duration prediction in bike-sharing context, which is more change-agostic than traditional station-level prediction methods
Employed a combination of geographic clustering and purpose-aware categorization of stations in bike-sharing context