Browse Teach Curate

Prune-and-Fuse Data Selection for Efficient Deep Network Training

PruneFuse: a training-efficiency method that prunes a dense network at initialization into a structurally coherent surrogate, uses that surrogate to run active-learning sample selection (Least Confidence, Entropy, or Greedy k-centers), then fuses the trained pruned model back into the dense model for a better initialization and refines it with knowledge distillation, cutting selection compute while matching state-of-the-art active-learning accuracy.

B9 Text

Pruning-based Data Selection and Network Fusion for Efficient Deep Learning Humaira Kousar∗ Hasnain

PruneFuse is a method that makes active-learning data selection and network training cheaper. A dense network is pruned at initialization (one-shot structured channel pruning by weight magnitude) int…