Scholarlink

<p>Effective hyperparameter optimization remains a significant challenge in deep learning, particularly when applied to fine-grained image classification tasks with high inter-class similarity. In this work, we propose a novel physics-inspired metaheuristic termed the String Theory Algorithm for tuning convolutional neural network hyperparameters. The method draws from theoretical string dynamics, modeling candidate configurations as oscillating entities that explore the search space through harmonic perturbations and probabilistic updates. We apply String Theory Algorithm to optimize key hyperparameters of the ResNet-50 architecture on a curated 500-class subset of the PlantCLEF dataset, a complex benchmark for species-level plant identification. Hyperparameters considered include learning rate, batch size, optimizer type, weight decay, layer freezing depth, augmentation strength, and dropout rate—spanning both continuous and categorical dimensions. The objective function, defined as Top-1 validation accuracy, is evaluated through full training cycles with early stopping to ensure resource efficiency. STA consistently outperforms conventional methods including grid search, random sampling, and more recent techniques like Bayesian Optimization, Hyperband, and Optuna. The best STA-tuned configuration achieves 92.4% Top-1 validation accuracy and 92.1% Top-1 test accuracy, exceeding other methods by a significant margin while requiring fewer evaluations. Comparative analysis also shows improvements in macro-level performance metrics such as precision, recall, F1-score, and AUC, alongside reduced variance across independent runs. These findings suggest that incorporating theoretical constructs from string physics into optimization routines can lead to both conceptual innovation and measurable gains in model performance. On a representative imbalanced PlantCLEF split, STA achieves Macro F1 = 68.4% and Balanced Accuracy = 74.9%, outperforming tuned Bayesian/Optuna baselines, highlighting robustness under real-world class skew.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing resnet-50 for PlantCLEF classification with a string theory-inspired hyperparameter search

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Sukanta Ghosh,
  • Vinod Kumar Shukla,
  • Amar Singh,
  • Jayanta Chanda

摘要

Effective hyperparameter optimization remains a significant challenge in deep learning, particularly when applied to fine-grained image classification tasks with high inter-class similarity. In this work, we propose a novel physics-inspired metaheuristic termed the String Theory Algorithm for tuning convolutional neural network hyperparameters. The method draws from theoretical string dynamics, modeling candidate configurations as oscillating entities that explore the search space through harmonic perturbations and probabilistic updates. We apply String Theory Algorithm to optimize key hyperparameters of the ResNet-50 architecture on a curated 500-class subset of the PlantCLEF dataset, a complex benchmark for species-level plant identification. Hyperparameters considered include learning rate, batch size, optimizer type, weight decay, layer freezing depth, augmentation strength, and dropout rate—spanning both continuous and categorical dimensions. The objective function, defined as Top-1 validation accuracy, is evaluated through full training cycles with early stopping to ensure resource efficiency. STA consistently outperforms conventional methods including grid search, random sampling, and more recent techniques like Bayesian Optimization, Hyperband, and Optuna. The best STA-tuned configuration achieves 92.4% Top-1 validation accuracy and 92.1% Top-1 test accuracy, exceeding other methods by a significant margin while requiring fewer evaluations. Comparative analysis also shows improvements in macro-level performance metrics such as precision, recall, F1-score, and AUC, alongside reduced variance across independent runs. These findings suggest that incorporating theoretical constructs from string physics into optimization routines can lead to both conceptual innovation and measurable gains in model performance. On a representative imbalanced PlantCLEF split, STA achieves Macro F1 = 68.4% and Balanced Accuracy = 74.9%, outperforming tuned Bayesian/Optuna baselines, highlighting robustness under real-world class skew.