Manhattan Rule for Robust In-Situ Training of Memristive Deep Neural Network Accelerators

Ellina Zhang, Jack Cai, Amirali Amirsoleimani, Mostafa Rahimi Azghadi, Roman Genov, Majid Ahmadi

Abstract: In this work, we propose an alternative training approach for memristive circuits - the Manhattan rule training - which utilizes only sign information for weight updates. We present an in-depth analysis in both in-situ and ex-situ settings and show that not only does our method simplify circuit design but it also improves neural network robustness against device non-idealities. Using the MemTorch and our custom in-situ training framework, we implemented the Manhattan rule for MNIST classification and ECG signal detection tasks and achieved close to state-of-the-art performance under noise. Our work also provides a thorough comparison of Manhattan and conventional training methods under the effects of various device non-idealities, giving a crucial benchmark useful for the design of biomedical neural circuits.

External IDs:dblp:conf/mwscas/ZhangCAAGA24