NN Algorithm Aware Alternate Layer Retraining on Computation-in-Memory for Write Variation Compensation of Non-volatile Memories at Edge AI

This study proposes an alternate retraining of neural network (NN) to recover inference accuracy degraded by non-volatile memory (NVM) write variation in Computation-in-Memory (CiM). The proposed alternate layer retraining minimizes write variation when rewriting NN weights to NVM in CiM by updating only error tolerant layers of NN models. The proposed retraining exploits the difference in error tolerance among NN layers, and achieves 23.4% higher accuracy recovery with fewer rewriting NN weights than conventional methods.