.png){kind=icon}
.png){kind=icon}
Return
MFKD: Multi-dimensional feature alignment for knowledge distillation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.imavis.2025.105514.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge distillation is a popular technique for compressing and transferring models in the field of deep learning. However, existing distillation methods often focus on optimizing a single dimension and overlook the importance of aligning and transforming knowledge across multiple dimensions, leading to suboptimal results. In this article, we introduce a novel approach called multi-dimensional feature alignment for knowledge distillation (MFKD) to address this limitation. The MFKD framework is built on the observation that knowledge from different dimensions can complement each other effectively. We extract knowledge from features in the spatcial, sample and channel dimensions separately. Our spatial-level part separates the foreground and background information, guiding the student to focus on crucial image regions by mimicking the teacher's spatial and channel attention maps. Our sample-level part distills knowledge encoded in semantic correlations between sample activations by aligning the student's activations to emulate the teacher's clustering patterns using the Spearman correlation coefficient. Furthermore, our channel-level part encourages the student to learn standardized feature representations aligned with the teacher's channel-wise interdependencies. Finally, we dynamically balance the loss factors of the different dimensions to optimize the overall performance of the distillation process. To validate the effectiveness of our methodology, we conduct experiments on benchmark datasets such as CIFAR-100, ImageNet and COCO. The experimental results demonstrate substantial performance improvements compared to baseline and recent state-of-the-art methods, confirming the efficacy of our MFKD framework. Furthermore, we provide a comprehensive analysis of the experimental results, offering deeper insight into the benefits and effectiveness of our approach. Through this analysis, we reinforce the significance of aligning and leveraging knowledge across multiple dimensions in knowledge distillation.
Keywords:
Knowledge distillation
Feature-based distillation
Journal
IF:
4.2
Papers:
4.0K
Citations:
6.7K