DIVER:Diving Deeper into Distilled Data via Expressive Semantic Recovery
Researchers have introduced DIVER, a novel dual-stage framework for dataset distillation that aims to improve privacy and learning efficiency. Unlike previous single-stage methods that overfit to specific architectures, DIVER uses a pre-trained diffusion model to recover and preserve intrinsic semantics. This approach enhances generalization across different architectures and requires significantly less GPU memory and processing time compared to existing methods. AI
IMPACT This new method for dataset distillation could lead to more efficient and private AI model training.