Learning both Weights and Connections for Efficient Neural Network
Introduces a three-step prune-and-retrain method that learns important connections to shrink networks by an order of magnitude with no accuracy loss.
Neural networks are compute- and memory-intensive, making them hard to deploy on embedded systems, and conventional training fixes the architecture beforehand. The authors cut storage and computation by an order of magnitude, without hurting accuracy, by learning only the important connections. The three-step method first trains the network to find important connections, then prunes the unimportant ones, and finally retrains to fine-tune the remaining weights. On ImageNet this cut AlexNet parameters 9x (61M to 6.7M) and VGG-16 13x (138M to 10.3M) with no accuracy loss.
Based on: Learning both Weights and Connections for Efficient Neural Network · Neural Information Processing Systems
Curated by Aramai Editorial
Read summary →