sdb

  Dr. Sreyasee Das Bhattacharjee, PhD
  Senior Member IEEE
  Associate Professor of Research & Teaching
  Davis Hall #349
  Department of Computer Science & Engg.
  State University of New York at Buffalo, NY 14260
  Ph: (716)-645-4769


Sreyasee Das Bhattacharjee is a faculty member in the Computer Science and Engineering department at SUNY Buffalo (UB). Before joining UB, she was an adjunct faculty and postdoctoral fellow at the Computer Science Dept of UNC Charlotte. Prior moving to the United States, she was a research fellow at Nanyang Technological University (NTU), Singapore. Sreyasee has worked as a Visiting Research Scholar at the University of Western Australia (UWA), Perth and scientist at the Aeronautical Development Lab (ADE) of Defence Research and Development Organization (DRDO), Govt of India, in Bangalore, India. She received the Ph.D. degree from Indian Institute of Technology, Madras, M.Tech. degree from Indian Institute of Technology, Delhi, and M. Sc Mathematics from Chennai Mathematical Institute, India.

Her current research interests cover multimodal data analytics, affective computing, educational data mining, natural scene understanding, data-centric generative models, multimodal quantum AI and generative systems, human-machine collaborations, and AI-powered immersive systems. She has collaborated with a number of external collaborators from JPL, NASA, Department of Astronomy, California Institute of Technology, Roswell Park Comprehensive Cancer Center, University of California at San Francisco, OMRON Industrial Automation, etc.

Update:

We are actively recruiting self-motivated and hard-working doctoral candidates, graduate and exceptional undergraduate students in the areas of Generative AI, Immersive Technology, Multisensor systems, and Multimodal Data Analytics. Some of the ongoing projects include:
  • Biosignal-based Affective State Analysis in the Wild
  • Multimodal Key Event Localization in Videos
  • BCI-AI integration for Decoding human thoughts in continuum
  • AI-powered Interactive Agents in VR
  • Quantum Machine Learning

News:

Present/Past Leadership Responsibilities :

Present/Past Services :

Fine-grained Human Motion Assessment via Task-Specific Kinematic Experts

Generic vision-language models recognize broad actions but miss the joint-level, phase-specific flaws (e.g., knee valgus) that determine exercise safety. This project develops a vision-language-kinematic framework for fine-grained human motion assessment and real-time posture coaching. It fuses video with 3D skeletal dynamics and uses region-aware anatomical masks to analyze the upper body, core, and lower body separately. A global motion encoder feeds an expanding Mixture-of-Experts, and a Gaussian Mahalanobis Expert Selection Mechanism routes each motion to its closest expert without task labels. New exercises are learned class-incrementally by warm-starting experts from kinematically similar ones while earlier experts stay frozen. On FLAG3D, HMI, and QEVD-FIT-COACH, the model reaches 98.61% Top-1 action recognition and the best zero-shot 3D pose error (223.30 mm MPJPE), improves motion-text alignment over prior motion-language models, and delivers well-timed coaching feedback.

Authors: Pruthvi Moluguri, Ritik Verma, Sreyasee Das Bhattacharjee, Junsong Yuan

Project Page: https://multimodal-human-sensing.github.io/motion-assessment/

Robust Multimodal Affective Computing: From Incomplete Sensing to Generative Biosignal Diffusion

Across this body of work, our team advances affective computing and multisensory emotion intelligence across wearable, conversational, and naturalistic environments, systematically overcoming persistent real-world constraints such as noisy sensors, missing modalities, class imbalance, and demographic bias. Our research program establishes robust multimodal representations and adaptive fusion frameworks—incorporating joint embeddings and knowledge distillation—to preserve predictive integrity despite incomplete or degraded sensor streams. Progressing from robust representation learning to generative architectures, our recent efforts pioneer conditional and atomic biosignal diffusion models coupled with chain-of-thought reasoning to synthesize minority patterns, mitigate systemic bias across heterogeneous populations, and scale the expressive reasoning capacity of physiological signals for dependable, fair, and real-time affective state assessment.

Authors/Project Contributors: Ritik Verma, Sreyasee Das Bhattacharjee, Sabir Thapa Chhetri, Wenyao Xu, Junsong Yuan, Pruthvi Moluguri

Project Page: https://multimodal-human-sensing.github.io/atomsignet/

Advancing Non-Invasive Brain-to-Image Synthesis via Multimodal Diffusion Architectures

Translating visual perception directly from non-invasive neural signals into high-fidelity imagery presents significant challenges due to signal noise, rapid temporal dynamics, and substantial inter-subject variability. Our research bridges this gap through an evolving generative decoding paradigm: BrainDiffNet first established the feasibility of cross-modal neural decoding by aligning non-invasive EEG representations with latent visual feature spaces to guide generative diffusion. Advancing this foundation, MIND-DiT introduces an end-to-end multimodal diffusion transformer that replaces decoupled pipelines with unified neural-visual modeling. By integrating a spatio-temporal encoder to capture channel-wise topography and temporal dynamics, an Adaptive Identity Evolution mechanism to continually personalize neural condition tokens for unseen subjects without full model retraining, and a Unified Neural-Visual Transformer that facilitates bidirectional cross-attention across all denoising steps, this progression establishes a scalable, adaptive framework for direct brain-to-image synthesis.

Authors/Project Contributors: Ritik Verma, Pruthvi Moluguri, Soumyadeep Bhattacharjee, Sreyasee Das Bhattacharjee

Project Page: https://multimodal-human-sensing.github.io/mind-dit/