Bonan Xu

Huiyuan Tian, Bonan Xu, Shijian Li

While feature-based knowledge distillation has proven highly effective for compressing CNNs, these techniques unexpectedly fail when applied to Vision Transformers (ViTs), often performing worse than simple logit-based distillation. We provide the first comprehensive analysis of this phenomenon through a novel analytical framework termed as "distillation dynamics", combining frequency spectrum analysis, information entropy metrics, and activation magnitude tracking. Our investigation reveals that ViTs exhibit a distinctive U-shaped information processing pattern: initial compression followed by expansion. We identify the root cause of negative transfer in feature distillation: a fundamental representational paradigm mismatch between teacher and student models. Through frequency-domain analysis, we show that teacher models employ distributed, high-dimensional encoding strategies in later layers that smaller student models cannot replicate due to limited channel capacity. This mismatch causes late-layer feature alignment to actively harm student performance. Our findings reveal that successful knowledge transfer in ViTs requires moving beyond naive feature mimicry to methods that respect these fundamental representational constraints, providing essential theoretical guidance for designing effective ViTs compression strategies. All source code and experimental logs are provided at https://github.com/thy960112/Distillation-Dynamics.

Huiyuan Tian, Bonan Xu, Shijian Li et al.

Knowledge Distillation (KD) has achieved widespread success in compressing large Vision Transformers (ViTs), but a unified theoretical framework for both ViTs and KD is still lacking. In this paper, we propose SpectralKD, a novel unified analytical framework that offers deeper insights into ViTs and optimizes KD via spectral analysis. Our model-wise analysis reveals that CaiT concentrates information in their first and last few layers, informing optimal layer selection for KD. Surprisingly, our layer-wise analysis discovers that Swin Transformer and CaiT exhibit similar spectral encoding patterns despite their architectural differences, leading to feature map alignment guideline. Building on these insights, we propose a simple yet effective spectral alignment method for KD. Benefiting from the deeper understanding by above analysis results, even such a simple strategy achieves state-of-the-art performance on ImageNet-1K without introducing any trainable parameters, improving DeiT-Tiny by $+5.2\%$ and Swin-Tiny by $+1.4\%$ in top-1 accuracy. Furthermore, our post-training analysis reveals that distilled students can reproduce spectral patterns similar to their teachers, opening a new area we term ``distillation dynamics". Code and experimental logs are available in https://github.com/thy960112/SpectralKD.

Huiyuan Tian, Bonan Xu, Shijian Li et al.

Feature-map knowledge distillation (KD) is highly effective for convolutional networks but often fails for Vision Transformers (ViTs). To understand this failure and guide method design, we conduct a two-view representation analysis of ViTs. First, a layer-wise Singular Value Decomposition (SVD) of full feature matrices shows that final-layer representations are globally low-rank: for CaiT-S24, only $121/61/34/14$ dimensions suffice to capture $99\%/95\%/90\%/80\%$ of the energy. In principle, this suggests that a compact student plus a simple linear projector should be enough for feature alignment, contradicting the weak empirical performance of standard feature KD. To resolve this paradox, we introduce a token-level Spectral Energy Pattern (SEP) analysis that measures how each token uses channel capacity. SEP reveals that, despite the global low-rank structure, individual tokens distribute energy over most channels, forming a high-bandwidth encoding pattern. This results in an encoding mismatch between wide teachers and narrow students. Motivated by this insight, we propose two minimal, mismatch-driven strategies: (1) post-hoc feature lifting with a lightweight projector retained during inference, or (2) native width alignment that widens only the student's last block to the teacher's width. On ImageNet-1K, these strategies reactivate simple feature-map distillation in ViTs, raising DeiT-Tiny accuracy from $74.86\%$ to $77.53\%$ and $78.23\%$ when distilling from CaiT-S24, while also improving standalone students trained without any teacher. Our analysis thus explains why ViT feature distillation fails and shows how exploiting low-rank structure yields effective, interpretable remedies and concrete design guidance for compact ViTs.