| Popis: |
Data augmentation, pivotal for enhancing model performance and generalization, has seen significant advancements with mixed-sample techniques that blend elements from multiple training instances. Despite their success, these methods face challenges like generating unrealistic samples and potential performance degradation with excessive use. Our study explores the application of traditional mixed-sample augmentations at the feature level within Convolutional Neural Networks (CNNs), acknowledging existing limitations and untapped potential. This exploration aims to understand how such an approach might enhance generalization by encouraging the model to learn more abstract representations. Drawing on insights from this investigation, we propose FFMix, a collection of four innovative feature-level augmentation methods. These methods employ region-specific augmentation strategies, including Channel Max Position Detector, which focuses on peak activation of each channel, and Top k Channel Patches Detector, which selects the most informative patches of each channel based on average values for varied augmentation. They also use advanced label mixing ratios determined by Feature Map Area Ratio, which calculates mixed labels based on feature area contributions within a mask, and Semantic Information Estimator, which assesses feature contributions by quantifying semantic information in each channel’s feature map, ensuring coherent augmentation. We conducted nine experiments across three tasks (generalization, fine-grained, and sound classification) using the FMMix technique on five datasets and models. Our approach tested four FMMix methods, comparing the best one with techniques like Mixup, Manifold Mixup, and CutMix and analyzing hyperparameters and loss landscapes. The best-performing FMMix method showed improved generalization and reliability over traditional methods, with a slight increase in resource use. Our source code and models are available at https://github.com/khoanta-ai/FMMix. |
