Extended Data Fig. 1: Trade-off dilemma of granularity in sparse training. | Nature Electronics

Extended Data Fig. 1: Trade-off dilemma of granularity in sparse training.

From: An index-free sparse neural network using two-dimensional semiconductor ferroelectric field-effect transistors

Extended Data Fig. 1

a, Illustration of process of sparsification and the granularity of sparse neural network (SPNN), layer-wise (the most coarse-grained), block-wise, vector-wise, and element-wise (the most fine-grained). b, The accuracy-efficiency trade-off dilemma of four granularities in sparse training. The data comes from the NeuroSim simulation of VGG8-Net on CIFAR-10 dataset, trained with IMC hardware (See more details in Methods and Extended Data Fig. 9).

Back to article page