Embedded Files
September 2026 - I am pleased to share that our paper "DCM-SAM: Defect-Conditioned Mixture of LoRA Experts for NPU-Deployed AM Defect Segmentation" has been accepted at the On-Device Intelligence: Foundation Models under Real-World Constraints workshop at NeurIPS 2026.
Building on our earlier XCT-SAM work, DCM-SAM adapts the Segment Anything Model to automatically detect two classes of defects in X-ray CT scans of metal additive manufacturing parts, training on synthetic data alone and updating only 4.4% of the model's parameters. A central contribution of this work goes beyond accuracy: we deploy the full adapted model on a Qualcomm Hexagon NPU at full image resolution, with every operator running on-device, demonstrating that foundation models can be made to run efficiently at the edge under real industrial constraints. Congratulations to my students and co-authors Md Mushfiqur Rahaman and Md Mahedi Hasan on this achievement. This work was supported by a grant from Qualcomm, whose continued backing of our manufacturing AI research we gratefully acknowledge.
September 2026 - Congratulations to my Ph.D. student Tasmiah Haque, in the Department of Industrial and Management Systems Engineering at WVU, whose work on vision-based fall detection has been accepted for presentation at the Women in Machine Learning (WiML) workshop at NeurIPS 2026. The work asks a deceptively simple question: what happens to vision-based fall detection when the body's landmarks are not fully visible to the camera? The research proposes a privacy-preserving fall-monitoring pipeline that replaces continuous video transmission with compact unsupervised motion keypoints, and shows that anatomical independence, meaning not being tied to predefined body joints, provides meaningful robustness advantages under occlusion and partial visibility, while dramatically reducing the compute needed for always-on edge deployment.
August 2026 - I am pleased to share that our paper "XCT-SAM: Sequential Parameter-Efficient Domain Adaptation of SAM for Industrial XCT Defect Segmentation" was presented at the IAPR Workshop on Machine Vision for Industrial Inspection (MVI2), held in conjunction with the 28th International Conference on Pattern Recognition (ICPR 2026) in Lyon, France.
The work addresses a hard problem in additive manufacturing: automatically detecting defects in X-ray CT scans of printed metal parts, where flaws appear as subtle microstructural anomalies and labeled data is scarce. Our approach adapts the Segment Anything Model to this specialized domain in stages, first learning from alloy microstructure images and then transferring to XCT data, to bridge the large gap between everyday photographs and industrial scans. Congratulations to my students and co-authors Md Mahedi Hasan and Md Mushfiqur Rahaman, and to our collaborators, on this achievement. This work was supported by a grant from Qualcomm, whose backing of our manufacturing AI research we gratefully acknowledge.
July 2026 - I had the privilege to give a talk on the history of artificial intelligence to Beni Madhab High School students at a small town in rural West Bengal, India; far from the hustle and bustle of the nearest megacity Kolkata. The room was full including boys and girls who had come to hear how machines learned to see, talk, and reason, from a summer workshop in 1956 to the systems of today.
What stayed with me was not the talk but the questions. How do the layers of a neural network actually work? How does a computer generate a face it has never seen? How does motion transfer work? These were not questions asked out of casual politeness to a guest lecturer. They were the questions of students who had followed the ideas closely and wanted to go further.
Many of them travel to school by bicycle and study in conditions with few resources. None of that dimmed the curiosity in the room. If the recent progress in AI has lowered the old barriers to entering the field, then students like these, with time, effort, and curiosity, are as well placed to shape what comes next as students anywhere else.
June 2026 - Congratulations to my Ph.D. student Tasmiah Haque, in the Department of Industrial and Management Systems Engineering at WVU, who successfully defended her doctoral proposal in June 2026 and advanced to Ph.D. candidacy. Her research develops generative time series models with applications to problems in healthcare and industrial engineering. Best wishes to Tasmiah as she moves toward completing her dissertation.
June 2026 - I am delighted to share that Bikash Kharel, a Ph.D. student in Astronomy at WVU, presented our work on representation learning for Fast Radio Bursts at the Physics and AI Conference (PAI 2026) at Stanford, a major gathering in the growing field of AI for Science. The work introduces MT-GMDN, a novel multimodal Transformer that infers the scattering timescale of Fast Radio Bursts directly from their observed burst morphology. By jointly learning from both the dynamic spectrum and the timeseries of a burst, the model replaces slow classical fitting pipelines with near-real-time probabilistic inference, while producing well-calibrated uncertainty estimates that traditional tools struggle to deliver at low signal-to-noise. Congratulations to Bikash on this accomplishment, and thanks to my collaborator Emmanuel Fonseca and the Astronomy group at WVU for a productive partnership at the intersection of AI and astrophysics. Details are as below:
May 2026 - My student Daniel Campa was named a 2026 National Science Foundation Graduate Research Fellow, one of four WVU students to earn this highly competitive award. Working with me in the School of Mathematical and Data Sciences, along with researchers from Advanced Micro Devices, Daniel co-authored our paper "Path Generation and Evaluation in Video Games: A Nonparametric Statistical Approach," accepted at the 2025 IEEE Conference on Games, which introduces a statistically grounded method for generating and evaluating human-like navigation paths in video games. Daniel graduated in May 2026 and will use the fellowship to pursue a Ph.D. in Robotics at Arizona State University. Congratulations to Daniel on receiving such a highly selective and prestigious fellowship, and best wishes to him on his future endeavors!
Page updated
Google Sites
Report abuse