2020 年 33 巻 5 号 p. 156-162
Neural networks have high performance in tasks such as image recognition. However, the computational cost is high, and it is difficult to implement them on small devices. In recent years, in order to solve this problem, research on compression techniques of the neural network has been advanced. “Pruning” is known as one of the important approaches for compression techniques. A “sensitivity map” is a map that visualizes which area of the input data the model focused on. However, it has not been analyzed much on how the model structual changes by pruning affects the sensitivity maps. In this paper, we analyze the influence of pruning on the sensitivity maps. As a result, it was found that the region of interest in the background of the sensitivity map was reduced after pruning.