Data-Free Knowledge Distillation with Soft Targeted Transfer Set Synthesis
This addresses the issue of data unavailability due to storage or privacy constraints for neural network compression, though it is incremental as it builds on existing knowledge distillation techniques.
The paper tackles the problem of knowledge distillation without access to the original training data by proposing a data-free method that models the teacher network's feature space with a multivariate normal distribution and uses soft targeted labels to synthesize pseudo samples, achieving competitive performance compared to using the original dataset and other data-free approaches.
Knowledge distillation (KD) has proved to be an effective approach for deep neural network compression, which learns a compact network (student) by transferring the knowledge from a pre-trained, over-parameterized network (teacher). In traditional KD, the transferred knowledge is usually obtained by feeding training samples to the teacher network to obtain the class probabilities. However, the original training dataset is not always available due to storage costs or privacy issues. In this study, we propose a novel data-free KD approach by modeling the intermediate feature space of the teacher with a multivariate normal distribution and leveraging the soft targeted labels generated by the distribution to synthesize pseudo samples as the transfer set. Several student networks trained with these synthesized transfer sets present competitive performance compared to the networks trained with the original training set and other data-free KD approaches.