Multi-Teacher Knowledge Distillation with Reinforcement Learning for Visual Recognition

Authors: Chuanguang Yang, XinQiang Yu, Han Yang, Zhulin An, Chengqing Yu, Libo Huang, Yongjun Xu

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on visual recognition tasks, including image classification, object detection, and semantic segmentation tasks, demonstrate that MTKD-RL achieves state-of-the-art performance compared to the existing multi-teacher KD works.
Researcher Affiliation Academia 1Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2University of Chinese Academy of Sciences, Beijing, China EMAIL
Pseudocode Yes Algorithm 1: Overall MTKD-RL Training Procedure Algorithm 2: Alternative Multi-Teacher KD and Agent Optimization
Open Source Code Yes Code https://github.com/winycg/MTKD-RL
Open Datasets Yes We use CIFAR-100 (Krizhevsky, Hinton et al. 2009) and Image Net (Deng et al. 2009) datasets for image classification experiments. The object detection experiments adopt COCO-2017 (Lin et al. 2014) dataset. We utilize Cityscapes (Cordts et al. 2016), ADE20K (Zhou et al. 2017) and COCO-Stuff-164K (Caesar, Uijlings, and Ferrari 2018) datasets for semantic segmentation.
Dataset Splits Yes We use CIFAR-100 (Krizhevsky, Hinton et al. 2009) and Image Net (Deng et al. 2009) datasets for image classification experiments. The object detection experiments adopt COCO-2017 (Lin et al. 2014) dataset. We utilize Cityscapes (Cordts et al. 2016), ADE20K (Zhou et al. 2017) and COCO-Stuff-164K (Caesar, Uijlings, and Ferrari 2018) datasets for semantic segmentation.
Hardware Specification Yes As shown in Table 6(c), we compare training complexity in one epoch with other methods over NVIDIA RTX 4090.
Software Dependencies No No specific software versions (e.g., Python, PyTorch, CUDA versions) are mentioned.
Experiment Setup Yes We set α = 1 and β = 5, as analysed in Fig.2. At first, we pre-train the student network S by multi-teacher KD following LMT KD (Equ.(2)) with equal weights, i.e. {wm i = 1}M m=1, for one training epoch.