Snehanshu Saha
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Snehanshu Saha has a total of 248 co-authors. Here are the top co-authors by number of shared publications:
- Vaskar Raychoudhury
Miami University
12 shared publications - Archana Mathur
Nitte Meenakshi Institute of Technology
10 shared publications Archana Mathur
Dept of Artificial Intelligence and Data Science, Nitte Meenakshi Institute of Technology, Nitte (Deemed to be University), Bangalore, 560064, Karnataka, India
9 shared publications- Nithin Nagaraj
Complex Systems Programme National Institute of Advanced Studies Indian Institute of Science Campus 560012 Bengaluru Karnataka India
8 shared publications - Soma S. Dhavala
Director-ML Wadhwani AI Bengaluru India
8 shared publications Jyotirmoy Sarkar
BITS Pilani, K. K. Birla Goa Campus, Goa, India
6 shared publications
Latest Publications
- 17 October 2026
- Published by Springer Nature
- p. 161-179
SurfaceEncoder: Semi-supervised Representation Learning for Global Wheelchair Accessibility
- 17 October 2026
- Published by Springer Nature
- p. 161-179
SurfaceEncoder: Semi-supervised Representation Learning for Global Wheelchair Accessibility
- Open Access
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 347-360
PowerQuant: Architecture-Agnostic GPU Power Estimation via Quantile Regression
Accurate prediction of NVIDIA GPU power consumption remains challenging due to rapid architectural evolution. Existing machine-learning–based power models are tightly coupled to specific GPU architectures and degrade sharply on unseen platforms, requiring retraining and extensive power measurements, which hinder scalability. This paper presents a quantile-regression–based GPU power prediction framework that enables architecture-agnostic power estimation using static analysis-based compile-time CUDA kernel features. The key insight is that architectural changes primarily induce systematic shifts in power scale, while the relative ordering of kernel power demands remains preserved. By learning power quantiles that capture this ordering and mapping them to new GPUs through one-time calibration, the proposed approach mitigates cross-architecture distribution shift. Extensive evaluation across multiple NVIDIA GPU generations shows that, on unseen architectures, the proposed method improves prediction accuracy by up to 30–50% over existing regression models, while maintaining comparable accuracy in in-distribution settings. The resulting low-overhead, generalizable power estimates make the approach practical for power-aware scheduling, energy budgeting, and sustainability-oriented resource management in large HPC systems.
- 13 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 347-360
Open AccessPowerQuant: Architecture-Agnostic GPU Power Estimation via Quantile Regression
Accurate prediction of NVIDIA GPU power consumption remains challenging due to rapid architectural evolution. Existing machine-learning–based power models are tightly coupled to specific GPU architectures and degrade sharply on unseen platforms, requiring retraining and extensive power measurements, which hinder scalability. This paper presents a quantile-regression–based GPU power prediction framework that enables architecture-agnostic power estimation using static analysis-based compile-time CUDA kernel features. The key insight is that architectural changes primarily induce systematic shifts in power scale, while the relative ordering of kernel power demands remains preserved. By learning power quantiles that capture this ordering and mapping them to new GPUs through one-time calibration, the proposed approach mitigates cross-architecture distribution shift. Extensive evaluation across multiple NVIDIA GPU generations shows that, on unseen architectures, the proposed method improves prediction accuracy by up to 30–50% over existing regression models, while maintaining comparable accuracy in in-distribution settings. The resulting low-overhead, generalizable power estimates make the approach practical for power-aware scheduling, energy budgeting, and sustainability-oriented resource management in large HPC systems.
- 20 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. PP (3) , 1-16
SLAB: A Self-supervised Label Generation Framework to Reduce Annotation Overhead
Recently, many emotion-aware services have been designed to improve the user experience during online video consumption (e.g., online meeting platforms, and online tutoring mediums). These services primarily deploy a machine learning model that infers user emotion continuously based on the variation in different physiological signals . However, the major challenge to developing such models is to continuously ask the users to self-report their emotions for ground truth annotation. To address these challenges, we present SLAB (Self-supervised LABel generator), a self-supervised deep learning framework for label generation to reduce the tabular data annotation overhead. SLAB utilizes two pretext tasks coupled with a novel $log-cosh$ loss function to learn useful representations from unlabeled data that apply to the downstream task of label generation. SLAB outperforms supervised learning, semi-supervised learning, and self-supervised learning baselines such as Value Imputation and Mask Estimation (VIME) on continuous emotion annotation datasets - Continuously Annotated Signals of Emotion (CASE) and K-EmoCon. We also demonstrate the generalizability of the SLAB on a unique wheelchair surface vibration dataset (in addition to four publicly available datasets from different domains). These findings highlight that SLAB significantly reduces the annotation overhead in situations where the number of unlabeled samples vastly outnumbers labeled samples.
- 20 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. PP (3) , 1-16
SLAB: A Self-supervised Label Generation Framework to Reduce Annotation Overhead
Recently, many emotion-aware services have been designed to improve the user experience during online video consumption (e.g., online meeting platforms, and online tutoring mediums). These services primarily deploy a machine learning model that infers user emotion continuously based on the variation in different physiological signals . However, the major challenge to developing such models is to continuously ask the users to self-report their emotions for ground truth annotation. To address these challenges, we present SLAB (Self-supervised LABel generator), a self-supervised deep learning framework for label generation to reduce the tabular data annotation overhead. SLAB utilizes two pretext tasks coupled with a novel $log-cosh$ loss function to learn useful representations from unlabeled data that apply to the downstream task of label generation. SLAB outperforms supervised learning, semi-supervised learning, and self-supervised learning baselines such as Value Imputation and Mask Estimation (VIME) on continuous emotion annotation datasets - Continuously Annotated Signals of Emotion (CASE) and K-EmoCon. We also demonstrate the generalizability of the SLAB on a unique wheelchair surface vibration dataset (in addition to four publicly available datasets from different domains). These findings highlight that SLAB significantly reduces the annotation overhead in situations where the number of unlabeled samples vastly outnumbers labeled samples.
- Open Access
Matching High-Dimensional Geometric Quantiles for Test-Time Adaptation of Transformers and Convolutional Networks Alike
Test-time adaptation (TTA) refers to adapting a classifier for the test data when the probability distribution of the test data slightly differs from that of the training data of the model. To the best of our knowledge, most of the existing TTA approaches modify the weights of the classifier relying heavily on the architecture. It is unclear as to how these approaches are extendable to generic architectures. In this article, we propose an architecture-agnostic approach to TTA by adding an adapter network pre-processing the input images suitable to the classifier. This adapter is trained using the proposed quantile loss. Unlike existing approaches, we correct for the distribution shift by matching high-dimensional geometric quantiles. We prove theoretically that under suitable conditions minimizing quantile loss can learn the optimal adapter. We validate our approach on CIFAR10-C, CIFAR100-C and TinyImageNet-C by training both classic convolutional and transformer networks on CIFAR10, CIFAR100 and TinyImageNet datasets.
Open AccessMatching High-Dimensional Geometric Quantiles for Test-Time Adaptation of Transformers and Convolutional Networks Alike
Test-time adaptation (TTA) refers to adapting a classifier for the test data when the probability distribution of the test data slightly differs from that of the training data of the model. To the best of our knowledge, most of the existing TTA approaches modify the weights of the classifier relying heavily on the architecture. It is unclear as to how these approaches are extendable to generic architectures. In this article, we propose an architecture-agnostic approach to TTA by adding an adapter network pre-processing the input images suitable to the classifier. This adapter is trained using the proposed quantile loss. Unlike existing approaches, we correct for the distribution shift by matching high-dimensional geometric quantiles. We prove theoretically that under suitable conditions minimizing quantile loss can learn the optimal adapter. We validate our approach on CIFAR10-C, CIFAR100-C and TinyImageNet-C by training both classic convolutional and transformer networks on CIFAR10, CIFAR100 and TinyImageNet datasets.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 131595-131610
QuikDel: A Scalable Hierarchical Multi-Agent System for Urban Delivery Using Q-Learning
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 131595-131610
Open AccessQuikDel: A Scalable Hierarchical Multi-Agent System for Urban Delivery Using Q-Learning
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 136953-136973
Identifying Surface Accessibility Through Vibration Classification: A Novel Utility Kernel Approach
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 136953-136973
Open AccessIdentifying Surface Accessibility Through Vibration Classification: A Novel Utility Kernel Approach
- Open Access
- 17 November 2025
- Published by Oxford University Press (OUP) in Monthly Notices of the Royal Astronomical Society
- Vol. 544 (4) , 4208-4216
Investigating the bulge morphology of dual AGN host galaxies from the gothic survey
We present a structural analysis of bulges in dual active galactic nuclei (AGNs) host galaxies. Dual AGNs arise in galaxy mergers where both supermassive black holes are actively accreting. The AGNs are typically embedded in compact bulges, which appear as luminous nuclei in optical images. Galaxy mergers can result in bulge growth, often via star formation. The bulges can be discy (pseudo-bulges), classical bulges, or belong to elliptical galaxies. Using Sloan Digital Sky Survey Data Release 18 gri images and galfit modelling, we performed 2D decomposition for 131 dual AGNs bulges (comprising 61 galaxy pairs and 3 galaxy triplets) identified in the gothic survey. We derived Sérsic indices, luminosities, masses, and scale lengths of the bulges. Most bulges (105/131) are classical, with Sérsic indices lying between $n=2$ and $n=8$. Among these, 64 per cent are elliptical galaxies, while the remainder are classical bulges in disc galaxies. Only ${\sim}20$ per cent of the sample exhibits pseudo-bulges. Bulge masses span $1.5\times 10^9 \, {\rm M}_\odot$ to $1.4\times 10^{12}\, {\rm M}_\odot$, with the most massive systems being ellipticals. Galaxy-type matching shows that elliptical–elliptical and elliptical–disc mergers dominate over disc–disc mergers. At least one galaxy in two-thirds of the dual AGN systems is elliptical and only ${\sim}30$ per cent involve two disc galaxies. Although our sample is limited, our results suggest that dual AGNs preferentially occur in evolved, red, quenched systems, which typically form via major mergers. They are predominantly hosted in classical bulges or elliptical galaxies rather than star-forming disc galaxies.- 17 November 2025
- Published by Oxford University Press (OUP) in Monthly Notices of the Royal Astronomical Society
- Vol. 544 (4) , 4208-4216
Open AccessInvestigating the bulge morphology of dual AGN host galaxies from the gothic survey
We present a structural analysis of bulges in dual active galactic nuclei (AGNs) host galaxies. Dual AGNs arise in galaxy mergers where both supermassive black holes are actively accreting. The AGNs are typically embedded in compact bulges, which appear as luminous nuclei in optical images. Galaxy mergers can result in bulge growth, often via star formation. The bulges can be discy (pseudo-bulges), classical bulges, or belong to elliptical galaxies. Using Sloan Digital Sky Survey Data Release 18 gri images and galfit modelling, we performed 2D decomposition for 131 dual AGNs bulges (comprising 61 galaxy pairs and 3 galaxy triplets) identified in the gothic survey. We derived Sérsic indices, luminosities, masses, and scale lengths of the bulges. Most bulges (105/131) are classical, with Sérsic indices lying between $n=2$ and $n=8$. Among these, 64 per cent are elliptical galaxies, while the remainder are classical bulges in disc galaxies. Only ${\sim}20$ per cent of the sample exhibits pseudo-bulges. Bulge masses span $1.5\times 10^9 \, {\rm M}_\odot$ to $1.4\times 10^{12}\, {\rm M}_\odot$, with the most massive systems being ellipticals. Galaxy-type matching shows that elliptical–elliptical and elliptical–disc mergers dominate over disc–disc mergers. At least one galaxy in two-thirds of the dual AGN systems is elliptical and only ${\sim}30$ per cent involve two disc galaxies. Although our sample is limited, our results suggest that dual AGNs preferentially occur in evolved, red, quenched systems, which typically form via major mergers. They are predominantly hosted in classical bulges or elliptical galaxies rather than star-forming disc galaxies. - Open Access
- 10 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 192492-192517
Non-Negative Matrix Factorization in Recommender Models: Concept, Survey, and Future Direction
Recommender Systems have gained attraction since last two decades owing to its popularity of providing customers with the information of their choice, items they might purchase, movies they might watch and many preferences. This has led recommenders to penetrate almost all e-commerce and industrial applications. The Netflix Prize competition has shown the rising path to recommenders and Singular Value Decomposition (SVD) based factorization techniques. SVD and its variants are considered as one of the efficient approach for recommendation task, especially aligned for collaborative filtering task. However, these SVD approaches remains less interpretable and explainable, due to the presence of negative components in latent features of entities. In a realistic scenario, it is not possible to have negative components in the features. This leads to exploration of factorization techniques focused on positive components only. The non-negative counterpart of SVD, Non-negative Matrix Factorization (NMF) imposes the non-negativity constraint to features, allowing for more interpretable results. It has wide applicability ranging from Bioinformatics, topic modeling to recommenders. Till date, there have been several survey dedicated to core working of NMF. However, there was very limited literature available as to how different NMF works with recommender systems. There was no streamlined survey on NMF based recommender models that explores NMF in depth with advanced domains, case study and future directions. This article presents the insight on NMF based recommender models with focus on their mathematical modeling and taxonomy based on information they use, penalization term and optimization approach. The article offers the NMF based recommender model based on emerging approaches like Deep Learning, Federated Learning, Transformers, Explainability and Dynamicity. The article also brings up the open issues and challenges of NMF models. Further, the article presented an experimental analysis on NMF models with SVD and Neural Network model. The experimental analysis was covered on three benchmarked datasets of recommenders along with stability analysis. The results showcased the performance improvement brought by NMF in recommenders. Through this article, one can plunge into the dynamics of NMF for recommenders.
- 10 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 192492-192517
Open AccessNon-Negative Matrix Factorization in Recommender Models: Concept, Survey, and Future Direction
Recommender Systems have gained attraction since last two decades owing to its popularity of providing customers with the information of their choice, items they might purchase, movies they might watch and many preferences. This has led recommenders to penetrate almost all e-commerce and industrial applications. The Netflix Prize competition has shown the rising path to recommenders and Singular Value Decomposition (SVD) based factorization techniques. SVD and its variants are considered as one of the efficient approach for recommendation task, especially aligned for collaborative filtering task. However, these SVD approaches remains less interpretable and explainable, due to the presence of negative components in latent features of entities. In a realistic scenario, it is not possible to have negative components in the features. This leads to exploration of factorization techniques focused on positive components only. The non-negative counterpart of SVD, Non-negative Matrix Factorization (NMF) imposes the non-negativity constraint to features, allowing for more interpretable results. It has wide applicability ranging from Bioinformatics, topic modeling to recommenders. Till date, there have been several survey dedicated to core working of NMF. However, there was very limited literature available as to how different NMF works with recommender systems. There was no streamlined survey on NMF based recommender models that explores NMF in depth with advanced domains, case study and future directions. This article presents the insight on NMF based recommender models with focus on their mathematical modeling and taxonomy based on information they use, penalization term and optimization approach. The article offers the NMF based recommender model based on emerging approaches like Deep Learning, Federated Learning, Transformers, Explainability and Dynamicity. The article also brings up the open issues and challenges of NMF models. Further, the article presented an experimental analysis on NMF models with SVD and Neural Network model. The experimental analysis was covered on three benchmarked datasets of recommenders along with stability analysis. The results showcased the performance improvement brought by NMF in recommenders. Through this article, one can plunge into the dynamics of NMF for recommenders.
- Open Access
- 1 November 2025
- Published by IOP Publishing in Publications of the Astronomical Society of the Pacific
- Vol. 137 (11) , 114502
Deconvolution for Large Astronomical Surveys: A Study of the Scaled Gradient Projection Method on Zwicky Transient Facility Data
Ground-based astronomical observations will continue to produce resolution-limited images due to atmospheric seeing. Deconvolution reverses such effects and thus can benefit extracted science in multifaceted ways. We apply the Scaled Gradient Projection algorithm for the single-band deconvolution of several observed images from the Zwicky Transient Facility and mainly discuss the performance on stellar sources. The method shows good photometric flux preservation, which deteriorates for fainter sources but significantly reduces flux uncertainties even for the faintest sources. Deconvolved sources have a well-defined Full-Width-at-Half-Maximum of roughly one pixel (one arcsecond for ZTF) regardless of the observed seeing. Detection after deconvolution results in catalogs with ≳99.6% completeness relative to detections in the observed images. A few observed sources that could not be detected in the deconvolved image are found near saturated sources, whereas for others, the deconvolved counterparts are detected when slightly different detection parameters are used. The deconvolution reveals new faint sources previously undetectable, which are confirmed by crossmatching with the deeper DESI Legacy DR10 and with Pan-STARRS1 through forced photometry. The method could identify examples of serendipitous potential deblends that exceeded SExtractor’s deblending capabilities, with as extreme as Δ m ≈ 3 and separations as small as one arcsecond between the deblended components. Our survey-agnostic approach is better and eight times faster than Richardson–Lucy deconvolution and could be a reliable method for incorporation into survey pipelines.
- 1 November 2025
- Published by IOP Publishing in Publications of the Astronomical Society of the Pacific
- Vol. 137 (11) , 114502
Open AccessDeconvolution for Large Astronomical Surveys: A Study of the Scaled Gradient Projection Method on Zwicky Transient Facility Data
Ground-based astronomical observations will continue to produce resolution-limited images due to atmospheric seeing. Deconvolution reverses such effects and thus can benefit extracted science in multifaceted ways. We apply the Scaled Gradient Projection algorithm for the single-band deconvolution of several observed images from the Zwicky Transient Facility and mainly discuss the performance on stellar sources. The method shows good photometric flux preservation, which deteriorates for fainter sources but significantly reduces flux uncertainties even for the faintest sources. Deconvolved sources have a well-defined Full-Width-at-Half-Maximum of roughly one pixel (one arcsecond for ZTF) regardless of the observed seeing. Detection after deconvolution results in catalogs with ≳99.6% completeness relative to detections in the observed images. A few observed sources that could not be detected in the deconvolved image are found near saturated sources, whereas for others, the deconvolved counterparts are detected when slightly different detection parameters are used. The deconvolution reveals new faint sources previously undetectable, which are confirmed by crossmatching with the deeper DESI Legacy DR10 and with Pan-STARRS1 through forced photometry. The method could identify examples of serendipitous potential deblends that exceeded SExtractor’s deblending capabilities, with as extreme as Δ m ≈ 3 and separations as small as one arcsecond between the deblended components. Our survey-agnostic approach is better and eight times faster than Richardson–Lucy deconvolution and could be a reliable method for incorporation into survey pipelines.