Progressively Knowledge Distillation via Re-parameterizing Diffusion Reverse Process

Abstract Knowledge distillation aims at transferring knowledge from the teacher model to the student one by aligning their distributions. Feature-level distillation often uses L2 distance or its variants as the loss function, based on the assumption that outputs follow normal distributions. This poses a significant challenge when distribution gaps are substantial since this loss function ignores the variance term. To address the problem, we propose to decompose the transfer objective into small parts and optimize it progressively. This process is inspired by diffusion models from which the noise distribution is mapped to the target distribution step by step. However, directly employing diffusion models is impractical in the distillation scenario due to its heavy reverse process. To overcome this challenge, we adopt the structural re-parameterization technique to generate multiple student features to approximate the teacher features sequentially. The multiple student features are combined linearly in inference time without extra cost. We present extensive experiments performed on various transfer scenarios, such as CNN-to-CNN and Transformer-to-CNN, that validate the effectiveness of our approach.