🤖 AI Summary
This work proposes an online data augmentation method grounded in control theory to address the limitations of handcrafted, non-adaptive augmentation strategies commonly used in visual recognition tasks. By constructing a control-loop architecture that leverages operation-specific response curves, the approach dynamically adjusts the intensity distribution of individual augmentation operations during training—eliminating the need for preset parameters and automatically suppressing augmentations detrimental to model performance. Experimental results demonstrate that, when applied with WideResNet-28-10 on CIFAR-10, CIFAR-100, and SVHN-core, the method achieves performance on par with state-of-the-art data augmentation techniques, offering a task-adaptive and fully automated dynamic augmentation mechanism without manual intervention.
📝 Abstract
We introduce ControlAugment (Ctrl-A), an automated data augmentation algorithm for image-vision tasks, which incorporates principles from control theory for online adjustment of augmentation strength distributions during model training. Ctrl-A eliminates the need for initialization of individual augmentation strengths. Instead, augmentation strength distributions are dynamically, and individually, adapted during training based on a control-loop architecture and what we define as relative operation response curves. Using an operation-dependent update procedure provides Ctrl-A with the potential to suppress augmentation styles that negatively impact model performance, alleviating the need for manually engineering augmentation policies for new image-vision tasks. Experiments on the CIFAR-10, CIFAR-100, and SVHN-core benchmark datasets using the common WideResNet-28-10 architecture demonstrate that Ctrl-A is highly competitive with existing state-of-the-art data augmentation strategies.