arXiv subjects

Ashutosh Sharma

Publications and source records attributed to Ashutosh Sharma.

At least 19 recordsLinked to original sources

arXiv2026

TileMaxSim: IO-Aware GPU MaxSim Scoring with Dimension Tiling and Fused Product Quantization

Multi-vector retrieval models such as ColBERT achieve state-of-the-art accuracy through fine-grained token-level MaxSim scoring, yet existing GPU implementations leave most hardware performance unused. We give a roofline analysis of MaxSim on modern GPUs and identify a severe bandwidth gap: naive implementations reach only 5-18% of peak HBM bandwidth because they materialize the Nq x Nd similarity matrix, wasting memory traffic on data that is consumed once and discarded. We present TileMaxSim, a family of IO-aware Triton kernels that close this gap via (1) multi-query SRAM tiling that streams document embeddings through shared memory while accumulating per-query-token maxima in registers, reading each embedding from HBM exactly once; (2) dimension tiling that partitions the embedding dimension into 128-wide chunks, enabling scoring for d > 128 embeddings that overflow shared memory; and (3) fused product-quantization scoring via shared-memory lookup tables, cutting HBM I/O by up to ~31x. On NVIDIA H100 GPUs, TileMaxSim reaches 80.2% of peak HBM bandwidth and scores 82M documents/second (71.6M/s on real MS MARCO passages), a 220x speedup over loop-based scoring, 6.5x over fused PyTorch, 6.6-8.5x over torch.compile, and 469x the scoring throughput of WARP's CPU engine on the same node. TileMaxSim preserves exact retrieval quality: on MS MARCO and three BEIR benchmarks, rankings match reference MaxSim. As a drop-in replacement in ColBERTv2/PLAID, it cuts scoring latency at 100K candidates from 268 ms to 1.2 ms (98% lower end-to-end latency). We further show constant throughput from 100K to 500K documents, data-parallel multi-GPU sharding, robustness across dimensions 64-768, and FP16/BF16/FP32 support. Concurrent work independently develops an IO-aware fused MaxSim kernel; we differ in dimension tiling for d > 128 and fused product-quantization scoring.

cs.IR↗

arXiv2026

GPUSparse: GPU-Accelerated Learned Sparse Retrieval with Parallel Inverted Indices

Learned sparse retrieval models such as SPLADE achieve retrieval quality competitive with dense models while preserving the interpretability and exact-match advantages of sparse representations. However, inference-time scoring still relies on CPU-bound inverted index traversal algorithms (WAND, Block-Max WAND), creating a fundamental bottleneck for real-time serving at scale. We present GPUSparse, a system for GPU-accelerated exact learned sparse retrieval that introduces: (1) a GPU-parallel inverted index with block-aligned, warp-coalesced posting lists; (2) a batched scatter-add scoring algorithm that processes hundreds of queries simultaneously; and (3) fused Triton kernels with an analysis of the tradeoff between work-efficiency and hardware utilization. On MS MARCO passage ranking (8.8M passages) with real SPLADE embeddings, GPUSparse matches CPU exact scoring to three decimals (MRR@10=0.383, equal to Pyserini SPLADE at this precision; Recall@1000>=0.999 vs. dense matmul, the residual from floating-point tie-breaking) while providing a 235x speedup over Pyserini CPU at 8.8M documents (1.27ms vs. 298ms per query). Compared to Seismic (the fastest CPU sparse retrieval system), which trades 25% recall for speed (R@1000=0.738 vs. 0.983 exact), GPUSparse achieves exact scoring at 787 QPS throughput (batch 500) on the full 8.8M collection, with 1.3ms per query. Our document-parallel kernel reaches 62.6% of H100 peak HBM bandwidth, revealing a fundamental work-efficiency vs. bandwidth-efficiency tradeoff in GPU sparse retrieval. The reformulation of sparse scoring as scatter-add over an inverted index is shared with SPARe's iterative mode; our contribution is its fused-kernel realization, which we measure to be 23-270x faster than a faithful SPARe iterative reimplementation.

cs.IR↗

arXiv2026

TractRLFusion: A GPT-Based Multi-Critic Policy Fusion Framework for Fiber Tractography

Tractography plays a pivotal role in the non-invasive reconstruction of white matter fiber pathways, providing vital information on brain connectivity and supporting precise neurosurgical planning. Although traditional methods relied mainly on classical deterministic and probabilistic approaches, recent progress has benefited from supervised deep learning (DL) and deep reinforcement learning (DRL) to improve tract reconstruction. A persistent challenge in tractography is accurately reconstructing white matter tracts while minimizing spurious connections. To address this, we propose TractRLFusion, a novel GPT-based policy fusion framework that integrates multiple RL policies through a data-driven fusion strategy. Our method employs a two-stage training data selection process for effective policy fusion, followed by a multi-critic fine-tuning phase to enhance robustness and generalization. Experiments on HCP, ISMRM, and TractoInferno datasets demonstrate that TractRLFusion outperforms individual RL policies as well as state-of-the-art classical and DRL methods in accuracy and anatomical reliability.

cs.LG↗

arXiv2025

INTHOP: A Second-Order Globally Convergent Method for Nonconvex Optimization

Second-order Newton-type algorithms that leverage the exact Hessian or its approximation are central to solve nonlinear optimization problems. However, their applications in solving large-scale nonconvex problems are hindered by three primary challenges: (1) the high computational cost associated with Hessian evaluations, (2) its inversion, and (3) ensuring descent direction at points where the Hessian becomes indefinite. We propose INTHOP, an interval Hessian-based optimization algorithm for nonconvex problems to deal with these primary challenges. The proposed search direction is based on approximating the original Hessian matrix by a positive definite matrix. The novelty of the proposed method is that the proposed search direction is guaranteed to be descent and requires approximation of Hessian and its inversion only at specific iterations. We prove that the difference between the calculated approximate and exact Hessian is bounded within an interval. Accordingly, the approximate Hessian matrix is reused if the iterates are in that chosen interval while computing the gradients at each iteration. We develop various algorithm variants based on the interval size updating methods and minimum eigenvalue computation methods. We also prove the global convergence of the proposed algorithm. Further, we apply the algorithm to an extensive set of test problems and compare its performance with the existing methods such as steepest descent, quasi-Newton, and Newton method. We show empirically that the proposed method solves more problems in fewer function and gradient evaluations than steepest descent and the quasi-Newton method. While in the comparison to the Newton method, we illustrate that for nonconvex optimization problems, we require substantially less $O(n^3)$ operations.

math.OC↗

arXiv2024

Tract-RLFormer: A Tract-Specific RL policy based Decoder-only Transformer Network

Fiber tractography is a cornerstone of neuroimaging, enabling the detailed mapping of the brain's white matter pathways through diffusion MRI. This is crucial for understanding brain connectivity and function, making it a valuable tool in neurological applications. Despite its importance, tractography faces challenges due to its complexity and susceptibility to false positives, misrepresenting vital pathways. To address these issues, recent strategies have shifted towards deep learning, utilizing supervised learning, which depends on precise ground truth, or reinforcement learning, which operates without it. In this work, we propose Tract-RLFormer, a network utilizing both supervised and reinforcement learning, in a two-stage policy refinement process that markedly improves the accuracy and generalizability across various data-sets. By employing a tract-specific approach, our network directly delineates the tracts of interest, bypassing the traditional segmentation process. Through rigorous validation on datasets such as TractoInferno, HCP, and ISMRM-2015, our methodology demonstrates a leap forward in tractography, showcasing its ability to accurately map the brain's white matter tracts.

cs.LG↗

arXiv2023

IndicIRSuite: Multilingual Dataset and Neural Information Models for Indian Languages

In this paper, we introduce Neural Information Retrieval resources for 11 widely spoken Indian Languages (Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu) from two major Indian language families (Indo-Aryan and Dravidian). These resources include (a) INDIC-MARCO, a multilingual version of the MSMARCO dataset in 11 Indian Languages created using Machine Translation, and (b) Indic-ColBERT, a collection of 11 distinct Monolingual Neural Information Retrieval models, each trained on one of the 11 languages in the INDIC-MARCO dataset. To the best of our knowledge, IndicIRSuite is the first attempt at building large-scale Neural Information Retrieval resources for a large number of Indian languages, and we hope that it will help accelerate research in Neural IR for Indian Languages. Experiments demonstrate that Indic-ColBERT achieves 47.47% improvement in the MRR@10 score averaged over the INDIC-MARCO baselines for all 11 Indian languages except Oriya, 12.26% improvement in the NDCG@10 score averaged over the MIRACL Bengali and Hindi Language baselines, and 20% improvement in the MRR@100 Score over the Mr.Tydi Bengali Language baseline. IndicIRSuite is available at https://github.com/saifulhaq95/IndicIRSuite

cs.IR↗

arXiv2023

DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines

The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded "prompt templates", i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, i.e. imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn (by creating and collecting demonstrations) how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric. We conduct two case studies, showing that succinct DSPy programs can express and optimize sophisticated LM pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, a few lines of DSPy allow GPT-3.5 and llama2-13b-chat to self-bootstrap pipelines that outperform standard few-shot prompting (generally by over 25% and 65%, respectively) and pipelines with expert-created demonstrations (by up to 5-46% and 16-40%, respectively). On top of that, DSPy programs compiled to open and relatively small LMs like 770M-parameter T5 and llama2-13b-chat are competitive with approaches that rely on expert-written prompt chains for proprietary GPT-3.5. DSPy is available at https://github.com/stanfordnlp/dspy

cs.CL↗

arXiv2023

Transforming Breast Cancer Diagnosis: Towards Real-Time Ultrasound to Mammogram Conversion for Cost-Effective Diagnosis

Ultrasound (US) imaging is better suited for intraoperative settings because it is real-time and more portable than other imaging techniques, such as mammography. However, US images are characterized by lower spatial resolution noise-like artifacts. This research aims to address these limitations by providing surgeons with mammogram-like image quality in real-time from noisy US images. Unlike previous approaches for improving US image quality that aim to reduce artifacts by treating them as (speckle noise), we recognize their value as informative wave interference pattern (WIP). To achieve this, we utilize the Stride software to numerically solve the forward model, generating ultrasound images from mammograms images by solving wave-equations. Additionally, we leverage the power of domain adaptation to enhance the realism of the simulated ultrasound images. Then, we utilize generative adversarial networks (GANs) to tackle the inverse problem of generating mammogram-quality images from ultrasound images. The resultant images have considerably more discernible details than the original US images.

eess.IV↗

arXiv2021

Critical Risk Indicators (CRIs) for the electric power grid: A survey and discussion of interconnected effects

The electric power grid is a critical societal resource connecting multiple infrastructural domains such as agriculture, transportation, and manufacturing. The electrical grid as an infrastructure is shaped by human activity and public policy in terms of demand and supply requirements. Further, the grid is subject to changes and stresses due to solar weather, climate, hydrology, and ecology. The emerging interconnected and complex network dependencies make such interactions increasingly dynamic causing potentially large swings, thus presenting new challenges to manage the coupled human-natural system. This paper provides a survey of models and methods that seek to explore the significant interconnected impact of the electric power grid and interdependent domains. We also provide relevant critical risk indicators (CRIs) across diverse domains that may influence electric power grid risks, including climate, ecology, hydrology, finance, space weather, and agriculture. We discuss the convergence of indicators from individual domains to explore possible systemic risk, i.e., holistic risk arising from cross-domains interconnections. Our study provides an important first step towards data-driven analysis and predictive modeling of risks in the coupled interconnected systems. Further, we propose a compositional approach to risk assessment that incorporates diverse domain expertise and information, data science, and computer science to identify domain-specific CRIs and their union in systemic risk indicators.

stat.AP↗

arXiv2018

High Energy electron and proton acceleration by circularly polarized laser pulse from near critical density hydrogen gas target

We demonstrate in this research the quasi-monoenergetic electron and proton acceleration through three dimensional particle-in-cell simulations of short petawatt circular polarized laser pulse interactions with near critical density hydrogen target. We numerically show that under controlled choice of laser and target parameters, the high energy electrons and protons can be illustrated in experiment at advanced high power laser facilities eg ELI - ALPS. We detailed the microphysics involved in the acceleration mechanism, which required investigating the role of plasma density gradients, plasma density, and target thickness. The role of selfgenerated plasma electric and magnetic fields is depicted on proton energy and density distribution. We numerically investigate here the laser driven proton acceleration where energetic protons with energies more than 200 MeV and charge in excess of 10 nC and conversion efficiency more than 6 percent (which implies 2.4 J proton beam out of the 40 J incident laser energy). Additionally and interestingly, we show from simulation study first time the quasi-monoenergetic ring shaped electron beam driven by circularly polarised laser which may prove useful for plasma based-based X-ray source and collimation of positron beam.

physics.plasm-ph↗

arXiv2012

Topological electronic structure and Weyl semimetal in the TlBiSe$_2$ class of semiconductors

We present an analysis of bulk and surface electronic structures of thallium based ternary III-V-VI$_2$ series of compounds TlMQ$_2$, where M=Bi or Sb and Q=S, Se or Te, using the ab initio density functional theory framework. Based on parity analysis and (111) surface electronic structure, we predict TlSbSe$_2$, TlSbTe$_2$, TlBiSe$_2$ and TlBiTe$_2$ to be non-trivial topological insulators with a single Dirac cone at the $Γ$-point, and TlSbS$_2$ and TlBiS$_2$ to be trivial band insulators. Our predicted topological phases agree well with available angle-resolved photoemission spectroscopy (ARPES) measurements, in particular the topological phase changes between TlBiSe$_2$ and TlBiS$_2$. Moreover, we propose that Weyl semimetal can be realized at the topological critical point in TlBi(S$_{1-x}$Se$_x$)$_2$ and TlBi(S$_{1-x}$Te$_x$)$_2$ alloys by breaking the inversion symmetry in the layer by layer growth in the order of Tl-Se(Te)-Bi-S, yielding six Dirac cones centered along the $Γ-L$ directions in the bulk band structure.

cond-mat.mes-hall↗

arXiv2012

Shear Mediated Elongational Flow and Yielding in Soft Glassy Materials

In this work, we study the deformation behavior of thin films of various soft glassy materials that are simultaneously subjected to two creep flow fields, rotational shear flow by applying torque and elongational flow by applying normal force. The generic behavior under the combined fields is investigated in different soft glassy materials with diverse microstructure such as: hair gel, emulsion paint, shaving foam and clay suspension. Increase in strength of one stress component while keeping the other constant, not only leads to an expected enhanced deformation in its own direction, but also greater strain in the other direction. Herschel Bulkley model is observed to explain this behavior qualitatively. Elongational flow induced in the materials eventually causes failure in the same. Interestingly time to failure is observed to be strongly dependent not just on normal force but also on the applied rotational shear stress. We believe that the presence of a three dimensional jammed structure, in which overall unjamming can be induced by applying stress having sufficient magnitude irrespective of the direction leads to the observed behavior. In addition, we observe self-similarity in the elongational as well as rotational strain time curves corresponding to various combinations of both the fields. This observation suggests a mere shift in the time-scales involved keeping the path followed in the process unchanged. A phase diagram is also constructed for various soft glassy materials by determining different combinations of orthogonal stresses beyond which materials yield. Estimated yield stress in the limit of flow dominated by applied tensile force on the top plate demonstrates scatter, which might be originating from fingering instability. Except this deviation, yielding is observed when the invariant of stress tensor exceeds yield stress, validating the Von Mises criterion.

cond-mat.soft↗

arXiv2011

Monte Carlo simulation of spin polarized transport in nanowires and 2-D channels of III-V semiconductors

We simulated spin polarized transport of electrons along III-V nanowires and two dimensional III-V channels using semi classical Monte Carlo method. Properties of spin relaxation length have been investigated in different III-V zinc-blende materials at various conditions, such as, temperature, external field etc. Spin dephasing in III-V channels is caused due to D'yakonov-Perel (DP) relaxation and due to Elliott-Yafet (EY) relaxation. Spin dephasing length in nanowire is found to be greater than that in 2-D channel.

cond-mat.mes-hall↗

arXiv2011

Monte Carlo simulation of spin relaxation in nanowires and 2-D channels of II-VI semiconductors

We have analysed spin relaxation behaviour of various II-VI semiconductors for nanowire structure and 2-D channel by simulating spin polarized transport through a semi-classical approach. Monte Carlo simulation method has been applied to simulate our model. D'yakanov-Perel mechanism and Elliot-Yafet mechanism are dominant for spin relaxation in II-VI semiconductors. Variation in spin relaxation length with external field has been analysed and comparison is drawn between nanowire and 2-D channels. Spin relaxation lengths of various II-VI semiconductors are compared at an external field of 1kV/cm to understand the predominant factors affecting spin de-phasing in them. Among the many results obtained, most noticeable one is that spin relaxation length in nanowires is many times greater than that in 2-D channel.

cond-mat.mes-hall↗

arXiv2011

Self-organized Nano-lens Arrays by Intensified Dewetting of Electron Beam Modified Polymer Thin-films

Sub-100 nm polymeric spherical plano-convex nano-lens arrays are fabricated using short electron beam exposures to selectively modify the ultrathin (< 30 nm) polymer films, followed by their intensified self-organized dewetting under an aqueous-organic mixture. A short exposure to e-beam locally modifies the polymer chains to effectively change the viscosity of the film in small domains, thus bringing in the dynamical dewetting contrast in the film that produces aligned and ordered dewetted nanostructures. Both negative and positive e-beam tone polymers are thus used to produce array of nano-lenses. The intensified self-organized dewetting under a water-organic solvent mixture overcomes the limitations on surface tension and dewetting force and thus facilitates the formation of sub-100 nm diameter polymer nanolenses of tunable curvature. By varying the extent of e-beam exposure, various configurations from isolated to connected nano-lens arrays can be fabricated.

cond-mat.soft↗

arXiv2011

Submicrometer Pattern Fabrication by Intensification of Instability in Ultrathin Polymer Films under a Water-Solvent Mix

Dewetting of ultrathin (< 100 nm) polymer films, by heating above the glass transition, produces droplets of sizes of the order of microns and mean separations between droplets of the order of tens of microns. These relatively large length scales are because of the weak destabilizing van der Waals forces and the high surface energy penalty required for deformations on small scales. We show a simple, one-step versatile method to fabricate sub-micron (>~100 nm) droplets and their ordered arrays by room temperature dewetting of ultrathin polystyrene (PS) films by minimizing these limitations. This is achieved by controlled room temperature dewetting under an optimal mixture of water, acetone and methyl-ethyl ketone (MEK). Diffusion of organic solvents in the film greatly reduces its glass transition temperature and the interfacial tension, but enhances the destabilizing field by introduction of electrostatic force. The latter is reflected in a change in the exponent, n of the instability length scale, λ ~h^n, where h is the film thickness and n = 1.51 \pm 0.06 in the case of water-solvent mix, as opposed to its value of 2.19 \pm 0.07 for dewetting in air. The net outcome is more than one order of magnitude reduction in the droplet size as well as their mean separation and also a much faster dynamics of dewetting. We also demonstrate the use of this technique for controlled dewetting on topographically patterned substrates with submicrometer features where dewetting in air is either arrested, incomplete or unable to produce ordered patterns.

cond-mat.soft↗

arXiv2011

Ultrafast Large Area Micropattern Generation in Non-absorbing Polymer Thin-Films by Pulsed Laser Diffraction

We report an ultrafast, parallel and beyond-the-master micro-patterning technique for ultrathin (30 nm-400 nm) non-absorbing polymer films by diffraction of a laser light through a two dimensional periodic aperture. The redistribution of laser energy absorbed by the substrate causes self-organization of polymer thin-film in the form of wrinkle like surface relief structures caused by localized melting and freezing of the thin-film. Unlike the conventional laser ablation and laser writing processes, low laser fluence is employed to only passively swell the polymer as a pre-ablative process without the loss of material, and without absorption/reaction with the incident radiation. Self-organization in the thin polymer film aided by the diffraction pattern produces micro-structures made up of thin raised lines. These regular microstructures have far more complex morphologies than the mask geometry and very narrow line widths that can be an order of magnitude smaller than the openings in the mask. The microstructure morphology is easily modulated by changing the film thickness, aperture size and geometry and by changing the diffraction pattern, e.g., by changing the aperture-substrate distance.

cond-mat.soft↗

arXiv2010

Enhanced Self-organized Dewetting of Ultrathin Polymer Films under Water-organic Solutions: Fabrication of Sub-micron Spherical Lens Arrays

Field-induced self-organized patterning in ultrathin (< 100 nm) polymer films produces resolutions of the order of 10 μm or more because of the high energy penalty for the surface deformations on small scales. We propose here a very simple but versatile method to fabricate sub-micron (~100 nm) ordered and tunable polymeric structures by self-organized room temperature dewetting of ultrathin polystyrene films by minimizing the surface tension limitation. We illustrate this technique by fabricating sub-micron lens arrays of tunable curvature. This is achieved by switching to controlled room temperature dewetting under an optimal mix of water, acetone and methyl-ethyl ketone (MEK). Organic solvents used decrease the glass transition temperature, greatly decrease the interfacial tension, intensify the field and increase the contact angle/aspect ratio of the resulting tunable nano-structures, without a concurrent solubilization of PS owing to water being the majority phase in the outside mixture.

cond-mat.soft↗