ReRAM-aware Model Finetuning addressing I-V Non-linearity and Retention Errors

In three linesHardware-aware finetuning method for DNN deployment on ReRAM crossbar arrays. Uses range-shrunk sinh transformation to mitigate I-V non-linearity and incorporates retention errors into regularization loss. Results: ResNet18/DeiT-Tiny no degradation, MobileNetV3 <2% on ImageNet, F-1 -1 point on SQuAD v2.

Summary generated by Claude — human-verified