Heterogeneous Learning Rates and Weight Fusion for Enhanced Image Classification Performance

이미지 분류 성능 향상을 위한 이종 학습률 및 가중치 융합 연구

This work proposes enhancing image classification performance through heterogeneous trajectory learning rates and weight fusion. Two models are independently trained using exploratory trajectory learning and local trajectory learning, then their weights are averaged to initialize re-adaptive optimization. The exploratory trajectory explores broad parameter space to avoid local optima, while the local trajectory ensures stable convergence. This weight fusion strategy leverages distinct optimization paths, guiding the model toward flatter loss regions that improve in-distribution test accuracy. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with ResNet, VGG, AlexNet, and ViT architectures show consistent improvements over baselines, with ResNet34 achieving a 2.92% gain on CIFAR-100. These findings demonstrate that this framework effectively enhances model performance without additional inference costs.