Behavioral Level Simulation Framework to Support Error-Aware CNN Training with In-Memory Computing

In recent years, in-memory computing (IMC) is a promising technique to solve the bottleneck of data movement in edge AI devices. To perform some simple computation in memory, the IMC designs often adopt analog operations, which may incur inevitable computation errors. To recover the accuracy loss, adding some random disturbance in CNN training is a straightforward approach to make it more tolerant to computation errors. However, random values may be quite different to the real run-time errors caused by non-ideal effects. In this work, we propose a hierarchical simulation framework for the IMC systems to support error-aware CNN training. This framework includes an efficient approach to build accurate IMC behavioral models that reflect real non-ideal effects. By using the behavioral model, an accurate high-level error model can be built efficiently to provide run-time errors for CNN training and error rate verification. As shown in the simulation results, the error-aware CNN training with the proposed models efficiently improves the CNN accuracy in real applications with almost no accuracy loss.