Knowledge distillation (KD) transfers the ``dark knowledge'' from a complex teacher model to a compact student model. However, heterogeneous architecture distillation, such as Vision Transformer (ViT) to ResNet18, faces challenges due to differences in spatial feature representations. Traditional KD methods are mostly designed for homogeneous architectures and hence struggle to effectively address the disparity. Although heterogeneous KD approaches have been developed recently to solve these issues, they often incur high computational costs and complex designs, or overly rely on logit alignment, which limits their ability to leverage the complementary features. To overcome these limitations, we propose Heterogeneous Complementary Distillation (HCD), a simple yet effective framework that integrates complementary teacher and student features to align representations in shared logits. These logits are decomposed and constrained to facilitate diverse knowledge transfer to the student. Specifically, HCD processes the student’s intermediate features through convolutional projector and adaptive pooling, concatenates them with teacher's feature from the penultimate layer and then maps them via the Complementary Feature Mapper (CFM) module, comprising fully connected layer, to produce shared logits. We further introduce Sub-logit Decoupled Distillation (SDD) that partitions the shared logits into n sub-logits, which are fused with teacher's logits to rectify classification. To ensure sub-logit diversity and reduce redundant knowledge transfer, we propose an Orthogonality Loss (OL). By preserving student-specific strengths and leveraging teacher knowledge, HCD enhances robustness and generalization in students. Extensive experiments on the CIFAR-100, fine-grained (e.g., CUB200, Aircraft) and ImageNet-1K datasets demonstrate that HCD outperforms state-of-the-art KD methods, establishing it as an effective solution for heterogeneous KD.

Knowledge distillation transfers "dark knowledge" from a large teacher model to a smaller student model, yielding a highly efficient network. To improve network's generalization ability, existing works use a larger temperature coefficient for knowledge distillation. Nevertheless, these methods may lower the target category's confidence and lead to ambiguous recognition of similar samples. To mitigate this issue, some studies introduce intra-batch distillation to reduce prediction discrepancy. However, these methods overlook the inconsistency between background information and the target category, which may increase prediction bias due to noise disturbance. Additionally, label imbalance from random sampling and batch size can undermine network generalization reliability. To tackle these challenges, we propose a simple yet effective Intra-class Knowledge Distillation (IKD) method that facilitates knowledge sharing within the same class to ensure consistent predictions. First, we initialize the matrix and the vector to store logits and class counts provided by the teacher, respectively. Then, in the first epoch, we calculate the sum of logits and sample counts per class and perform KD to prevent knowledge omission. Finally, in subsequent training, we update the matrix to obtain the average logits and compute the KL divergence between the student's output and the updated matrix according to the label index. This process ensures intra-class consistency and improves the student's performance. Furthermore, this method theoretically reduces prediction bias by ensuring intra-class consistency. Extensive experiments on the CIFAR-100, ImageNet-1K, and Tiny-ImageNet datasets validate the superiority of IKD.

State-of-the-art logit distillation methods exhibit versatility, simplicity, and efficiency. Despite the advances, existing studies have yet to delve thoroughly into fine-grained relationships within logit knowledge. In this paper, we propose Local Dense Relational Logit Distillation (LDRLD), a novel method that captures inter-class relationships through recursively decoupling and recombining logit information, thereby providing more detailed and clearer insights for student learning. To further optimize the performance, we introduce an Adaptive Decay Weight (ADW) strategy, which can dynamically adjust the weights for critical category pairs using Inverse Rank Weighting (IRW) and Exponential Rank Decay (ERD). Specifically, IRW assigns weights inversely proportional to the rank differences between pairs, while ERD adaptively controls weight decay based on total ranking scores of category pairs. Furthermore, after the recursive decoupling, we distill the remaining non-target knowledge to ensure knowledge completeness and enhance performance. Ultimately, our method improves the student's performance by transferring fine-grained knowledge and emphasizing the most critical relationships. Extensive experiments on datasets such as CIFAR-100, ImageNet-1K, and Tiny-ImageNet demonstrate that our method compares favorably with state-of-the-art logit-based distillation approaches. The code will be made publicly available.

Knowledge distillation (KD) is a technique that transfers “dark knowledge” from a deep teacher network (teacher) to a shallow student network (student). Despite significant advances in KD, existing work has not adequately mined two crucial types of knowledge: 1) the knowledge of head categories, which represents the relationship between the target category and its similar categories. Our findings reveal that this highly similar (complex) knowledge is essential for improving student’s performance; and 2) the effectively utilized knowledge of tail categories. Existing studies often treat the non-target categories collectively without sufficiently considering the effectiveness of knowledge from tail categories. To tackle these challenges, we reformulate classical KD (ReKD) into two components: Top-KInter-class Similar Distillation (TISD) and Non-Top-KInter-class Discriminability (NTID). Firstly, TISD captures and imparts the knowledge of head categories to the student. Our experimental results have verified that TISD is particularly effective in transferring the knowledge of head categories, even in fine-grained dataset classification. Secondly, we theoretically show that the weighting coefficient of NTID increases with the probability of Top-K, leading to stronger suppression of knowledge transfer for tail categories. This observation explains why difficult samples are more informative than simple ones. To better utilize both types of knowledge, we optimize both TISD and NTID using different weighting coefficients, thereby enhancing the student’s ability to learn this valuable knowledge from both head and tail categories. Furthermore, our extensive experimental results demonstrate that ReKD achieves state-of-the-art performance on various image classification datasets, including CIFAR-100, Tiny-ImageNet, and ImageNet-1K, as well as object detection and instance segmentation using the MS-COCO dataset.

Knowledge distillation is a deep learning method that mimics the way that humans teach, i.e., a teacher network is used to guide the training of a student one. Knowledge distillation can generate an efficient student network to facilitate deployment in resource-constrained edge computing devices. Existing studies have typically mined knowledge from a teacher network and transferred it to a student one. The latter can only passively receive knowledge but cannot understand how the former acquires the knowledge, thus limiting the latter’s performance improvement. Inspired by the old Chinese saying “Give a man a fish and you feed him for a day; teach a man how to fish and you feed him for a lifetime,” this work proposes a Skill-transferring Knowledge Distillation (SKD) method to boost a student network’s ability to create new valuable knowledge. SKD consists of two main meta-learning networks: Teacher Behavior Teaching and Teacher Experience Teaching. The former captures the process of a teacher network’s learning behavior in the hidden layers and can predict the teacher network’s subsequent behavior based on previous ones. The latter models the optimal empirical knowledge of a teacher network’s output layer at each learning stage. With their help, a teacher network can provide its actions to a student one in the subsequent behavior and its optimal empirical knowledge in the current stage. SKD’s performance is verified through its application to multiple object recognition tasks and comparison with the state of the art.