Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FL-NAS: Towards Fairness of NAS for Resource Constrained Devices via Large Language Models

Ruiyang Qin, Yuting Hu, Zheyu Yan, Jinjun Xiong, Ahmed Abbasi, Yiyu Shi

TL;DR

FL-NAS tackles the problem of designing fair, resource-efficient DNNs for constrained devices by leveraging Large Language Models to guide Neural Architecture Search. The framework couples an LLM-based prompt generator, an LLM-based design generator, and a DNN evaluator to optimize three objectives: accuracy, fairness, and hardware efficiency, with fairness metrics (Unfairness, EODD, EOPP1, EOPP2) embedded in the evaluation loop. Empirical results on the ISIC2019 dermatology dataset show FL-NAS achieves higher accuracy (up to 11.4%), substantially better fairness (up to 18x improvements in unfairness and related metrics), and reduced model size, memory footprint, and latency (up to 4x, 18x, and 55x respectively) compared to baselines. The findings indicate LLMs can reason about nuanced design considerations beyond accuracy and enable faster NAS cycles (several GPU-hours versus thousands), pointing to practical routes for deploying fair, efficient models on edge devices. Overall, FL-NAS demonstrates a promising direction for fairness-aware NAS driven by foundation models, warranting further exploration and refinement.”

Abstract

Neural Architecture Search (NAS) has become the de fecto tools in the industry in automating the design of deep neural networks for various applications, especially those driven by mobile and edge devices with limited computing resources. The emerging large language models (LLMs), due to their prowess, have also been incorporated into NAS recently and show some promising results. This paper conducts further exploration in this direction by considering three important design metrics simultaneously, i.e., model accuracy, fairness, and hardware deployment efficiency. We propose a novel LLM-based NAS framework, FL-NAS, in this paper, and show experimentally that FL-NAS can indeed find high-performing DNNs, beating state-of-the-art DNN models by orders-of-magnitude across almost all design considerations.

FL-NAS: Towards Fairness of NAS for Resource Constrained Devices via Large Language Models

TL;DR

FL-NAS tackles the problem of designing fair, resource-efficient DNNs for constrained devices by leveraging Large Language Models to guide Neural Architecture Search. The framework couples an LLM-based prompt generator, an LLM-based design generator, and a DNN evaluator to optimize three objectives: accuracy, fairness, and hardware efficiency, with fairness metrics (Unfairness, EODD, EOPP1, EOPP2) embedded in the evaluation loop. Empirical results on the ISIC2019 dermatology dataset show FL-NAS achieves higher accuracy (up to 11.4%), substantially better fairness (up to 18x improvements in unfairness and related metrics), and reduced model size, memory footprint, and latency (up to 4x, 18x, and 55x respectively) compared to baselines. The findings indicate LLMs can reason about nuanced design considerations beyond accuracy and enable faster NAS cycles (several GPU-hours versus thousands), pointing to practical routes for deploying fair, efficient models on edge devices. Overall, FL-NAS demonstrates a promising direction for fairness-aware NAS driven by foundation models, warranting further exploration and refinement.”

Abstract

Neural Architecture Search (NAS) has become the de fecto tools in the industry in automating the design of deep neural networks for various applications, especially those driven by mobile and edge devices with limited computing resources. The emerging large language models (LLMs), due to their prowess, have also been incorporated into NAS recently and show some promising results. This paper conducts further exploration in this direction by considering three important design metrics simultaneously, i.e., model accuracy, fairness, and hardware deployment efficiency. We propose a novel LLM-based NAS framework, FL-NAS, in this paper, and show experimentally that FL-NAS can indeed find high-performing DNNs, beating state-of-the-art DNN models by orders-of-magnitude across almost all design considerations.
Paper Structure (15 sections, 4 equations, 1 figure, 1 table, 2 algorithms)

This paper contains 15 sections, 4 equations, 1 figure, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background and Motivations
  3. Neural Architecture Search
  4. NAS under Resource and Fairness Constraints
  5. NAS via Large Language Models
  6. FL-NAS Design Details
  7. Fairness Metrics for DNNs
  8. FL-NAS Framework Design
  9. LLM-based Prompt Generator
  10. LLM-based Design Generator
  11. DNN Evaluator
  12. Experimental Results
  13. Experimental Setup
  14. Evaluations
  15. Conclusion

Figures (1)

  • Figure 1: An example of LLM outputs related to DNN.