Advances in Knowledge Distillation and Multi-Task Learning

The field of knowledge distillation and multi-task learning is rapidly evolving, with a focus on improving the efficiency and effectiveness of model training. Recent developments have explored the use of dynamic balancing parameters, multimodal distillation, and relative feature enhanced meta-learning to address challenges such as real-time performance and class imbalance. These innovations have led to significant improvements in model compression, transfer learning, and predictive power. Notably, researchers have proposed novel frameworks that integrate multiple modalities, such as caption-guided supervision and object-centric masking, to enhance dataset distillation. Additionally, adaptive multi-task distillation methods have been developed to jointly model multiple tasks, reducing storage and training requirements. Some particularly noteworthy papers include: REMEDI, which achieved significant improvements in imbalanced prediction; MoKD, which introduced a multi-objective optimization framework for knowledge distillation; and JointDistill, which proposed an adaptive multi-task distillation method for joint depth estimation and scene segmentation.

Sources

Importance Analysis for Dynamic Control of Balancing Parameter in a Simple Knowledge Distillation Setting

Robust Understanding of Human-Robot Social Interactions through Multimodal Distillation

REMEDI: Relative Feature Enhanced Meta-Learning with Distillation for Imbalanced Prediction

MoKD: Multi-Task Optimization for Knowledge Distillation

Leveraging Multi-Modal Information to Enhance Dataset Distillation

JointDistill: Adaptive Multi-Task Distillation for Joint Depth Estimation and Scene Segmentation

Built with on top of