Srinjoy Das, Ph.D.
Srinjoy Das, Ph.D.
Assistant Professor of Data Science, School of Mathematical and Data Sciences
West Virginia University
Embedded Files
I am a Tenure-Track Assistant Professor in the School of Mathematical and Data Sciences (SMDS) at West Virginia University (WVU), where I have been a faculty member since Fall 2021. I also hold adjunct appointments in the Department of Industrial and Management Systems Engineering (IMSE), the Lane Department of Computer Science and Electrical Engineering (LCSEE), and the Department of Mechanical and Aerospace Engineering (MMAE), and am a member of the Center for Gravitational Waves and Cosmology (GWAC) in the Department of Physics and Astronomy. Prior to joining WVU, I was a Postdoctoral Researcher in the Department of Mathematics at UC San Diego. I received my Ph.D. in Electrical Engineering and M.S. in Statistics from UC San Diego, an M.S. in Electrical and Computer Engineering from UC Irvine, and a dual degree in Physics and Electrical Engineering from BITS Pilani, India.
My research develops statistically rigorous AI methods for time series analysis, generative modeling, and foundation models, with a unifying focus on deployability in resource-constrained and data-limited settings. These methods have been applied across diverse domains including climate analysis, telehealth, manufacturing, and astronomy. Before transitioning to academia, I spent over a decade in the semiconductor industry at Rockwell, Qualcomm, and Atheros, where I architected and led ASIC implementations of wireless baseband and location subsystems, resulting in issued patents in low-power clock design and high-throughput bus architecture, with multiple devices entering volume production. I maintain active funded collaborations with industry partners including AMD, Intel, and Qualcomm, offering students unique opportunities at the intersection of cutting-edge AI research and real-world applications.
Academic Background
Academic Background
Ph.D in Electrical Engineering (2018)
University of California, San Diego M.S. in Statistics (2018)
University of California, San Diego Postdoctoral Research Scholar (2019-2021)
University of California, San Diego - Department of MathematicsM.S. in Electrical and Computer Engineering (2005)
University of California, IrvineM.Sc (Honors) in Physics (1995)
Birla Institute of Technology and Science, PilaniB.E. (Honors) in Electrical and Electronics Engineering (1995)
Birla Institute of Technology and Science, Pilani
University of California, San Diego M.S. in Statistics (2018)
University of California, San Diego Postdoctoral Research Scholar (2019-2021)
University of California, San Diego - Department of MathematicsM.S. in Electrical and Computer Engineering (2005)
University of California, IrvineM.Sc (Honors) in Physics (1995)
Birla Institute of Technology and Science, PilaniB.E. (Honors) in Electrical and Electronics Engineering (1995)
Birla Institute of Technology and Science, Pilani
Research Interests
Research Interests
Statistical Methods for Forecasting and Synthetic Data Generation in Low-Data Regimes
Statistical Methods for Forecasting and Synthetic Data Generation in Low-Data Regimes
My research develops nonparametric and statistically principled methods for forecasting, clustering, and synthetic data generation in settings where data are limited or nonstationary. A core focus is predictive inference for locally stationary time series and random fields, where I develop model-free techniques that construct point predictors and uncertainty estimates without restrictive parametric assumptions. For clustering, I develop kernel-based distance measures that account for partial and non-overlapping matches across time series instances, with demonstrated advantages over Euclidean and dynamic time warping based alternatives. These methods have been applied to climate science, healthcare, and synthetic data generation for autonomous agent training, with a consistent emphasis on distributional reasoning, interpretability, and statistical validity.
My research develops nonparametric and statistically principled methods for forecasting, clustering, and synthetic data generation in settings where data are limited or nonstationary. A core focus is predictive inference for locally stationary time series and random fields, where I develop model-free techniques that construct point predictors and uncertainty estimates without restrictive parametric assumptions. For clustering, I develop kernel-based distance measures that account for partial and non-overlapping matches across time series instances, with demonstrated advantages over Euclidean and dynamic time warping based alternatives. These methods have been applied to climate science, healthcare, and synthetic data generation for autonomous agent training, with a consistent emphasis on distributional reasoning, interpretability, and statistical validity.
Representation Learning with Generative and Foundation Models
Representation Learning with Generative and Foundation Models
My research develops and adapts generative and foundation models to extract structured, low-dimensional representations for inference in domains where data are scarce, costly to label, or subject to stringent deployment constraints. In video-based applications, I develop bandwidth-efficient motion transfer pipelines that represent complex motion through sparse keypoint representations combined with generative time series models, enabling high-fidelity video synthesis under severe network constraints for telehealth and immersive media. For manufacturing, I apply foundation models such as the Segment Anything Model to perform unsupervised defect segmentation in additive manufacturing, using novel prompt generation and domain adaptation strategies that eliminate the need for labeled training data. Extending these methods to scientific discovery, I develop representation-learning approaches for problems in astronomy where physically structured, label-scarce data demand statistically principled inference. Across these efforts, my work couples statistical rigor with scalable generative and foundation modeling to achieve reliable representation and inference under limited data and real-world constraints.
My research develops and adapts generative and foundation models to extract structured, low-dimensional representations for inference in domains where data are scarce, costly to label, or subject to stringent deployment constraints. In video-based applications, I develop bandwidth-efficient motion transfer pipelines that represent complex motion through sparse keypoint representations combined with generative time series models, enabling high-fidelity video synthesis under severe network constraints for telehealth and immersive media. For manufacturing, I apply foundation models such as the Segment Anything Model to perform unsupervised defect segmentation in additive manufacturing, using novel prompt generation and domain adaptation strategies that eliminate the need for labeled training data. Extending these methods to scientific discovery, I develop representation-learning approaches for problems in astronomy where physically structured, label-scarce data demand statistically principled inference. Across these efforts, my work couples statistical rigor with scalable generative and foundation modeling to achieve reliable representation and inference under limited data and real-world constraints.
Inference Optimization and Hardware-Aware Algorithms
Inference Optimization and Hardware-Aware Algorithms
A central goal of my research is enabling state-of-the-art generative and foundation models to operate efficiently on resource-constrained edge platforms, driven by the recognition that real-world deployment demands low-latency, energy-efficient inference. My early work developed efficient deconvolution algorithms that minimize redundant memory accesses and enable energy-efficient image upsampling at the edge, establishing a principled framework for hardware-aware algorithm design. Building on this foundation, I have developed layerwise pruning methods for generative and discriminative deep neural networks, and more recently have investigated post-training quantization of sequence-based architectures and foundation models, yielding new insights into the role of calibration data composition and class balance in achieving reliable low-bit inference under realistic edge deployment conditions. Ongoing work extends this framework to the quantization and edge deployment of large-scale vision foundation models in collaboration with industry partners.
A central goal of my research is enabling state-of-the-art generative and foundation models to operate efficiently on resource-constrained edge platforms, driven by the recognition that real-world deployment demands low-latency, energy-efficient inference. My early work developed efficient deconvolution algorithms that minimize redundant memory accesses and enable energy-efficient image upsampling at the edge, establishing a principled framework for hardware-aware algorithm design. Building on this foundation, I have developed layerwise pruning methods for generative and discriminative deep neural networks, and more recently have investigated post-training quantization of sequence-based architectures and foundation models, yielding new insights into the role of calibration data composition and class balance in achieving reliable low-bit inference under realistic edge deployment conditions. Ongoing work extends this framework to the quantization and edge deployment of large-scale vision foundation models in collaboration with industry partners.
Page updated
Google Sites
Report abuse