Adaptive Training Controller (ATC): An Autotuning Framework for CNN-Based Plastic Waste Classification Using the PLASTIC Dataset

Technical Summary: Adaptive Training Controller (ATC) for CNN-Based Plastic Waste Classification

Problem Statement
Plastic waste pollution remains a critical ecological challenge, necessitating precise and automated sorting solutions. While deep learning-based image classification has shown promise, its application in waste management is hindered by two primary factors: the scarcity of domain-specific benchmark datasets and the lack of generalized, transparent autotuning algorithms. Existing approaches often rely on general garbage datasets that lack sufficient plastic-specific categories and visual variability, leading to poor model generalization. Furthermore, traditional hyperparameter optimization methods, such as Optuna, often require architecture-specific adjustments and multiple training trials, resulting in high computational costs and GPU dependency without guaranteeing consistent performance improvements across different Convolutional Neural Network (CNN) architectures.

Methodology
The research addresses these challenges through a dual-pronged approach involving dataset creation and a novel training framework.

1. The PLASTIC Dataset: The authors constructed a custom benchmark dataset named "PLASTIC," comprising 4,500 images distributed across 15 distinct classes of plastic waste (e.g., cosmetic plastics, medical plastics, Styrofoam, PVC). The dataset was curated from various open-source repositories and commercial stock image sites to ensure diversity in shape, texture, color, and surface deterioration. The data was split into 70% training, 20% validation, and 10% testing. Preprocessing involved resizing images to 224x224, normalization using ImageNet statistics, and data augmentation techniques including random resizing, flipping, rotation, and color jittering.

2. Model Evaluation: Ten popular CNN architectures were rigorously benchmarked under identical experimental conditions: ResNet (18, 34, 50, 101, 152), VGG (16, 19), GoogLeNet, EfficientNet, and AlexNet.

3. Training Strategies: The models were evaluated using three distinct training strategies:

   • Baseline: Fixed hyperparameters (SGD optimizer, constant learning rate of 0.001, no autotuning).
   • Optuna Autotuning: A Bayesian hyperparameter search framework used to optimize parameters prior to training.
   • Adaptive Training Controller (ATC): A proposed, architecture-agnostic autotuning framework. Unlike static search methods, ATC operates as a closed-loop, training-aware system. It dynamically adjusts the training process in real-time based on loss trends. Key mechanisms include:
     • Loss-Sensitive Learning Rate Adaptation: Uses an Exponential Moving Average (EMA) of the loss to detect stagnation. If stagnation is detected, the learning rate is decayed.
     • Scheduling: Implements cosine annealing with warm-up.
     • Stability Measures: Includes gradient clipping and Automatic Mixed Precision (AMP) for hardware-aware resource management.
     • Optimizer Flexibility: Supports SGD, Adam, and AdamW.

Key Contributions

• PLASTIC Dataset: A unique benchmark of 4,500 real-life images across 15 plastic waste categories, designed to overcome the limitations of general garbage datasets.
• Adaptive Training Controller (ATC): A generalized, transparent, and model-agnostic autotuning algorithm. It improves convergence and reduces computational costs by dynamically regulating learning rates, utilizing mixed precision, and stabilizing gradients without requiring architecture-specific tuning or extensive GPU resources.
• Comparative Analysis: A comprehensive evaluation of ten CNN architectures on both the new PLASTIC dataset and a standard public garbage dataset, demonstrating the superiority of domain-specific data and the ATC framework over baseline and Optuna methods.

Experimental Results
The study was conducted on a system equipped with an NVIDIA GeForce GTX 1660 Ti GPU.

• Performance on PLASTIC Dataset: ATC consistently yielded the highest accuracy values compared to both baseline and Optuna methods.

  • ResNet101 achieved the peak accuracy of 94.89% using ATC (compared to 94.44% with Optuna and 91.78% with baseline).
  • ResNet50 also achieved 94.89% accuracy with ATC, significantly outperforming its Optuna result of 92.89%.
  • ATC demonstrated superior stability, particularly for architectures like VGG16 and VGG19, where Optuna resulted in significant performance degradation (e.g., VGG19 test accuracy dropped to 77.78% with Optuna but remained at 89.33% with ATC).
  • ATC reduced training time and memory requirements while maintaining high accuracy.

• Performance on Garbage Dataset: When tested on the general garbage dataset, ATC again outperformed baseline and Optuna, though absolute accuracy was lower due to dataset heterogeneity. For instance, EfficientNet achieved 68.95% accuracy with ATC, compared to 66.78% with Optuna. Notably, ATC provided substantial gains for models that struggled with the baseline, such as ResNet50, which improved from 21.27% (baseline) to 55.44% (ATC).

• Efficiency: ATC required only a single training run with dynamic adjustments, whereas Optuna required multiple independent trials (n_trials), resulting in significantly higher computational overhead and GPU dependency for Optuna.

Significance and Claims
The paper claims that the proposed ATC framework offers a fast, effective, and computationally efficient solution for deep learning-based plastic waste classification. By integrating adaptive learning rate tuning, intelligent optimization, and hardware awareness, ATC minimizes the reliance on high-end GPUs and reduces training time while maximizing classification accuracy. The study asserts that the combination of the domain-specific PLASTIC dataset and the architecture-independent ATC framework provides a robust, scalable solution for sustainable waste management, outperforming traditional static training and search-based hyperparameter optimization methods across diverse CNN models. The authors note that while ATC shows significant improvement for lightweight and mid-depth models, deeper models may show marginal gains due to architectural saturation relative to the dataset size. Future work is directed toward developing lightweight models for edge computing and expanding the PLASTIC dataset.