arXiv subjects

Utkarsh Anand

Publications and source records attributed to Utkarsh Anand.

7 recordsLinked to original sources

arXiv2026

Distributional Decomposition of Consumption Inequality Change During COVID-19

We decompose the U.S. consumption inequality distributional changes during the COVID-19 phase. Analyzing the Consumption Expenditure Interview Survey data, we decompose observed changes in consumption inequality into components attributable to several individual variables. Using a distribution regression method, we construct counterfactual distributions under the scenario in which the consumption structure or any specific variable would have remained the same between the two years before and after the onset of the COVID-19 pandemic. We find that changes in the conditional distribution of consumption explain most of the observed decline in consumption inequality among male-headed households between 2018 and 2022. The rise in asset holdings has significantly increased the consumption inequality in all measures. Moreover, the changes in a set of household characteristics have significantly reduced the consumption inequality. Our analyses focus on the well-measured consumption components that are robust to the measurement errors in consumption data.

econ.GN↗

arXiv2025

Zorse: Optimizing LLM Training Efficiency on Heterogeneous GPU Clusters

Large language models (LLMs) require vast amounts of GPU compute to train, but limited availability and high costs of GPUs make homogeneous clusters impractical for many organizations. Instead, assembling heterogeneous clusters by pooling together GPUs of different generations allows them to achieve higher aggregate compute and make use of all available GPUs. However, training on heterogeneous clusters presents several challenges, including load balancing across GPUs, optimizing memory usage to accommodate varying memory capacities, and ensuring communication-efficient training over diverse network interconnects potentially spanning multiple datacenters. In this paper, we make the case that efficient training on heterogeneous clusters requires (1) the integration of pipeline parallelism and data parallelism in a manner that is both communication- and memory-efficient, and (2) a more adaptable configuration of pipeline and data parallelism, which includes the capability to flexibly partition GPUs into asymmetric pipeline parallel stages and to incorporate heterogeneous GPUs within the same data parallelism group. We propose Zorse, the first system to unify all these capabilities while incorporating a planner that automatically configures training strategies for a given workload. Our evaluation shows that Zorse significantly outperforms state-of-the-art systems in heterogeneous training scenarios.

cs.DC↗

arXiv2025

ASR-FAIRBENCH: Measuring and Benchmarking Equity Across Speech Recognition Systems

Automatic Speech Recognition (ASR) systems have become ubiquitous in everyday applications, yet significant disparities in performance across diverse demographic groups persist. In this work, we introduce the ASR-FAIRBENCH leaderboard which is designed to assess both the accuracy and equity of ASR models in real-time. Leveraging the Meta's Fair-Speech dataset, which captures diverse demographic characteristics, we employ a mixed-effects Poisson regression model to derive an overall fairness score. This score is integrated with traditional metrics like Word Error Rate (WER) to compute the Fairness Adjusted ASR Score (FAAS), providing a comprehensive evaluation framework. Our approach reveals significant performance disparities in SOTA ASR models across demographic groups and offers a benchmark to drive the development of more inclusive ASR technologies.

cs.SD↗

arXiv2024

Design and control of a robotic payload stabilization mechanism for rocket flights

The use of parallel manipulators in aerospace engineering has gained significant attention due to their ability to provide improved stability and precision. This paper presents the design, control, and analysis of 'STEWIE', which is a three-degree-of-freedom (DoF) parallel manipulator robot developed by members of the thrustMIT rocketry team, as a payload stabilization mechanism for their sounding rocket, 'Altair'. The goal of the robot was to demonstrate the attitude control of the parallel plate against the continuous change in orientation experienced by the rocket during its flight, stabilizing the payloads. At the same time, the high gravitational forces (G-forces) and vibrations experienced by the sounding rocket are counteracted. A novel design of the mechanism, inspired by a standard Stewart platform, is proposed which was down-scaled to fit inside a 4U CubeSat within its space constraints. The robot uses three micro servo motors to actuate the links that control the alignment of the parallel plate. In addition to the actuation mechanism, a robust control system for its manipulation was developed for the robot. The robot represents a significant advancement in the field of space robotics in the aerospace industry by demonstrating the successful implementation of complex robotic mechanisms in small, confined spaces such as CubeSats, which are standard form factors for large payloads in the aerospace industry.

cs.RO↗

arXiv2024

Cephalo: Harnessing Heterogeneous GPU Clusters for Training Transformer Models

Training transformer models requires substantial GPU compute and memory resources. In homogeneous clusters, distributed strategies allocate resources evenly, but this approach is inefficient for heterogeneous clusters, where GPUs differ in power and memory. As high-end GPUs are costly and limited in availability, heterogeneous clusters with diverse GPU types are becoming more common. Existing methods attempt to balance compute across GPUs based on capacity but often underutilize compute due to memory constraints. We present Cephalo, a system that optimizes compute and memory usage by decoupling compute distribution from training state assignment. Cephalo outperforms state-of-the-art methods by achieving significantly higher training throughput while supporting larger models and batch sizes.

cs.DC↗

arXiv2023

Non-Pyrotechnic Radial Deployment Mechanism for Payloads in Sounding Rockets

A novel, non-pyrotechnic payload deployment mechanism tailored for sounding rockets is introduced in this research paper. The mechanism addresses the challenge of efficiently and compactly deploying payloads radially during a single launch, featuring a cylindrical carrier structure actuated by a rack-pinion mechanism. Powered by a servo motor, the carrier structure translates to enable radial ejection of payloads. The paper presents the mechanism's design and conducts a comprehensive performance analysis, including structural stability, system dynamics and power requirements. A simulation model is developed to assess payload deployment behavior under various conditions, demonstrating the mechanism's viability and efficiency for deploying multiple payloads within a single sounding rocket launch. The mechanism's adaptability to accommodate diverse payload types, sizes and weights enhances its versatility, while its radial deployment capability allows payloads to be released at different altitudes, offering greater flexibility for scientific experiments. The paper concludes that this innovative payload radial deployment mechanism represents a significant advancement in sounding rocket technology and holds promise for a wide array of applications in both scientific and commercial missions.

eess.SY↗

arXiv2023

Optimization of a Runge-Kutta 4th Order Method-based Airbrake Control System for High-Speed Vehicles Using Neural Networks

The Runge-Kutta 4th Order (RK4) technique is extensively employed in the numerical solution of differential equations for airbrake control system design. However, its computational efficacy may encounter restrictions when dealing with high-speed vehicles that experience intricate aerodynamic forces. Using a Neural Network, a unique technique to improving the RK4-based airbrakes code is provided. The Neural Network is trained on numerous aspects of the high-speed vehicle as well as the current status of the airbrakes. This data was generated through the traditional RK4-based simulations and can predict the state of the airbrakes for any given state of the rocket in real-time. The proposed approach is demonstrated on a high-speed airbrakes control system, achieving comparable or better performance than the traditional RK4-based system while significantly reducing computational time by reducing the number of mathematical operations. The proposed method can adapt to changes in flow conditions and optimize the airbrakes system in real-time.

math.OC↗