Finetuning Method Improves DNN Deployment on ReRAM In-Memory Computing.

Traditional computing architectures face performance bottlenecks, leading to increased interest in In-Memory Computing (IMC) using ReRAM crossbar arrays for their density and energy efficiency. However, practical deployment of ReRAM is hindered by hardware non-idealities like I-V non-linearity and retention errors. Existing hardware-aware training methods often require extensive retraining, which is impractical for large, modern AI models. To overcome this, researchers have developed a finetuning-based hardware-aware training algorithm. This approach allows for robust deployment of deep neural networks on ReRAM with significantly reduced training overhead. It addresses I-V non-linearity through a range-shrunk sinh transformation and integrates retention errors directly into a regularization loss during the finetuning process. Evaluations across various models and tasks, including image classification and question-answering, demonstrate the method's effectiveness. It maintains accuracy comparable to base models for ResNet18 and DeiT-Tiny, shows less than 2% accuracy degradation for MobileNetV3 on ImageNet, and only a 1-point F-1 score degradation on SQuAD v2, proving its viability for large-scale AI deployment on ReRAM.