Self-Distillation for Randomized Neural Networks

Abstract

Knowledge distillation (KD) is a conventional method in the field of deep learning that enables the transfer of dark knowledge from a teacher model to a student model, consequently improving the performance of the student model. In randomized neural networks, due to the simple topology of network architecture and the insignificant relationship between model performance and model size, KD is not able to improve model performance. In this work, we propose a self-distillation pipeline for randomized neural networks: the predictions of the network itself are regarded as the additional target, which are mixed with the weighted original target as a distillation target containing dark knowledge to supervise the training of the model. All the predictions during multi-generation self-distillation process can be integrated by a multi-teacher method. By induction, we have additionally arrived at the methods for infinite self-distillation (ISD) of randomized neural networks. We then provide relevant theoretical analysis about the self-distillation method for randomized neural networks. Furthermore, we demonstrated the effectiveness of the proposed method in practical applications on several benchmark datasets.