ApproxTuner 2.0: Towards quality-driven approximation tuning

Rambhia, Vidhi

Permalink

https://hdl.handle.net/2142/129683 Copy

Description

Title

ApproxTuner 2.0: Towards quality-driven approximation tuning

Author(s)

Rambhia, Vidhi

Issue Date

2025-04-10

Director of Research (if dissertation) or Advisor (if thesis)

Adve, Vikram

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Approximate Computing
• Approximation Tuning
• Edge Computing
• Model Optimization
• Model Merging

Abstract

Deploying deep neural networks in real-world applications often requires a balance between model fidelity and resource efficiency. Traditional approximation techniques are usually applied in isolation and evaluated using proxy metrics such as accuracy, which may not reflect actual downstream task performance. This work presents ApproxTuner 2.0 building on top of ApproxTuner [1] and [2], a system for application-aware approximation tuning that puts the downstream task at the center of the optimization process. This system explores a configuration space of approximations using modular, pluggable components—knobs, applications, and QoS evaluators—and scores each configuration using domain-specific metrics that directly reflect utility. We validate our approach across two case studies, monocular depth estimation with DepthAnythingV2 and object tracking with YOLOv8, demonstrating that ApproxTuner can uncover configurations with 4× speedups without compromising application-level quality. Complementing this, we also discuss LEWIS (LayEr-WIse Sparsity) [3], a guided model merging technique that approximates traditional fine-tuning by combining task vectors from pre-trained models. Rather than relying on expensive retraining or naive averaging, LEWIS uses layerwise-activation norm deltas to guide sparsity during model merging. This offers a practical approximation of fine-tuning, especially in resource-constrained scenarios. Our experiments demonstrate that LEWIS significantly improves model merging effectiveness. Together, ApproxTuner and LEWIS represent two complementary axes of approximate computing for real-world AI: one focuses on tuning approximation strategies for a given model and a downstream task, and the other facilitates rapid adaptation across tasks by approximating fine-tuning itself.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129683

Copyright and License Information

Copyright 2025 Vidhi Rambhia

Owning Collections