[22p-E102-1~14] FS.1 Focused Session "AI Electronics"

[22p-E102-7] Neural Network Accuracy Evaluation Simulator for Computation-in-Memory with Non ideality of Memory Devices

〇Kazuhide Higuchi¹, Chihiro Matsui¹, Naoko Misawa¹, Ken Takeuchi¹ (1.Univ. of Tokyo)

Keywords:Computation-in-Memory, Non Volatile Memory, Neural Network

Computation-in-Memory (CiM) exploits memory array structure to achieve multiply-and-accumulate (MAC) operation that most consumes computation resources in deep neural networks (DNNs). The proposed accuracy evaluation simulator can reproduce the non-ideality of the device in CiM memory cells by arbitrarily quantizing the weights in the convolutional and fully connected layers, adding variability to the weights according to an arbitrary distribution, and adding and subtracting a constant value.