Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MAT-Agent: Adaptive Multi-Agent Training Optimization

Jusheng Zhang, Kaitong Cai, Yijia Fan, Ningyuan Liu, Keze Wang

TL;DR

MAT-Agent reframes multi-label image classification training as a dynamic, multi-agent decision process where four autonomous agents continuously adapt augmentation, optimization, learning-rate scheduling, and loss functions. Guided by a shared non-stationary bandit reward that balances accuracy, stability, and convergence, the framework achieves state-of-the-art performance and faster convergence across Pascal VOC, COCO, and VG-256, with strong cross-domain generalization. Comprehensive experiments and ablations demonstrate the necessity of inter-agent coordination and the value of a composite reward, including robustness to long-tail distributions. The approach offers a scalable, search-free alternative to traditional AutoML and hyperparameter optimization methods for complex, evolving training landscapes.

Abstract

Multi-label image classification demands adaptive training strategies to navigate complex, evolving visual-semantic landscapes, yet conventional methods rely on static configurations that falter in dynamic settings. We propose MAT-Agent, a novel multi-agent framework that reimagines training as a collaborative, real-time optimization process. By deploying autonomous agents to dynamically tune data augmentation, optimizers, learning rates, and loss functions, MAT-Agent leverages non-stationary multi-armed bandit algorithms to balance exploration and exploitation, guided by a composite reward harmonizing accuracy, rare-class performance, and training stability. Enhanced with dual-rate exponential moving average smoothing and mixed-precision training, it ensures robustness and efficiency. Extensive experiments across Pascal VOC, COCO, and VG-256 demonstrate MAT-Agent's superiority: it achieves an mAP of 97.4 (vs. 96.2 for PAT-T), OF1 of 92.3, and CF1 of 91.4 on Pascal VOC; an mAP of 92.8 (vs. 92.0 for HSQ-CvN), OF1 of 88.2, and CF1 of 87.1 on COCO; and an mAP of 60.9, OF1 of 70.8, and CF1 of 61.1 on VG-256. With accelerated convergence and robust cross-domain generalization, MAT-Agent offers a scalable, intelligent solution for optimizing complex visual models, paving the way for adaptive deep learning advancements.

MAT-Agent: Adaptive Multi-Agent Training Optimization

TL;DR

MAT-Agent reframes multi-label image classification training as a dynamic, multi-agent decision process where four autonomous agents continuously adapt augmentation, optimization, learning-rate scheduling, and loss functions. Guided by a shared non-stationary bandit reward that balances accuracy, stability, and convergence, the framework achieves state-of-the-art performance and faster convergence across Pascal VOC, COCO, and VG-256, with strong cross-domain generalization. Comprehensive experiments and ablations demonstrate the necessity of inter-agent coordination and the value of a composite reward, including robustness to long-tail distributions. The approach offers a scalable, search-free alternative to traditional AutoML and hyperparameter optimization methods for complex, evolving training landscapes.

Abstract

Multi-label image classification demands adaptive training strategies to navigate complex, evolving visual-semantic landscapes, yet conventional methods rely on static configurations that falter in dynamic settings. We propose MAT-Agent, a novel multi-agent framework that reimagines training as a collaborative, real-time optimization process. By deploying autonomous agents to dynamically tune data augmentation, optimizers, learning rates, and loss functions, MAT-Agent leverages non-stationary multi-armed bandit algorithms to balance exploration and exploitation, guided by a composite reward harmonizing accuracy, rare-class performance, and training stability. Enhanced with dual-rate exponential moving average smoothing and mixed-precision training, it ensures robustness and efficiency. Extensive experiments across Pascal VOC, COCO, and VG-256 demonstrate MAT-Agent's superiority: it achieves an mAP of 97.4 (vs. 96.2 for PAT-T), OF1 of 92.3, and CF1 of 91.4 on Pascal VOC; an mAP of 92.8 (vs. 92.0 for HSQ-CvN), OF1 of 88.2, and CF1 of 87.1 on COCO; and an mAP of 60.9, OF1 of 70.8, and CF1 of 61.1 on VG-256. With accelerated convergence and robust cross-domain generalization, MAT-Agent offers a scalable, intelligent solution for optimizing complex visual models, paving the way for adaptive deep learning advancements.
Paper Structure (53 sections, 33 equations, 10 figures, 12 tables)

This paper contains 53 sections, 33 equations, 10 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Methodology
  4. Sequential Decision Formulation for MLIC Training Optimization
  5. MAT-Agent: Framework, Actions, and State Representation
  6. Decision-Making and Learning of Decentralized Adaptive Agents
  7. Dynamic Training Configuration Generation and MLIC Model Update
  8. Experiments
  9. Comparative Experiments: Multi-Dataset Evaluation
  10. Training Convergence Analysis
  11. Cross-dataset Generalization Ability.
  12. Analysis of Differences between Different Domains.
  13. Quantitative Comparisons with Existing Automated Methods.
  14. Ablation Study
  15. Conclusion
  16. ...and 38 more sections

Figures (10)

  • Figure 1: Framework of MAT-Agent: a multi-agent system that dynamically selects training strategies (augmentation, optimizer, scheduler, loss) based on current and historical training states to optimize multi-label classification.
  • Figure 2: Training loss (a) and mAP (b) changes for MAT-Agent and three baseline models. MAT-Agent shows faster convergence.
  • Figure 3: The distribution of policy attention weights of the MAT-Agent on different datasets
  • Figure 4: The relationship between the mAP and Rare-F1 of the MAT-Agent and mainstream baseline models over time during training
  • Figure 5: Ablation results on Pascal VOC. Removing any component degrades performance, while the full MAT-Agent achieves the best results across all metrics.
  • ...and 5 more figures