Image Aesthetics Assessment with Gradient Shake Regularization

Due to their large capacity, the convolutional networks are susceptible to prioritize memorization of the training examples over the problem understanding and generalization. This phenomenon is encountered mainly in the difficult problems, where classes are overlapping due to various reasons, including noisy annotations. In this paper we propose to regularize the training process of the convolutional neural networks by the injection of randomized perturbation within controlled amplitude of gradient of the loss function. The method is evaluated on aesthetic evaluation of images. The proposed solution is shown to improve the performance of baselines and to compare favorably with established strong solutions.