Single pulse blind-write with gradient accumulation strategy tolerant to non-idealities of memristive synapse for in-memory learning

Abstract Memristor-based analog in-memory learning (AIML) has emerged as a promising approach to improve energy efficiency in deep neural network training. However, non-idealities in memristive devices, such as nonlinearity, asymmetry, and cycle-to-cycle (C2C) and device-to-device (D2D) variations, pose significant challenges. These issues lead to increased energy consumption, reduced write precision, and compromised in-situ learning performance. To address these problems, we propose a mixed-precision training strategy that combines gradient accumulation with single pulse blind write method. We analyze the failure mechanisms of in-situ learning without these techniques and systematically investigate how various non-idealities affect AIML performance. We demonstrate that, by using our GA-Single Pulse strategy, high accuracy (95.36%) can be achieved even under significant non-idealities, including device conductance states being limited to 10 pulses for potentiation as well as 5 pulses for depression, the asymmetry of conductance state constrained to a factor of 2, the nonlinearity in long-term potentiation/ long-term depression curve reaching up to 5, C2C variation as high as 50%, and D2D variation extending up to 40% for learning handwritten digits in MNIST handwritten digit dataset, outperforms all previous reports. The results suggest that the idealities of memristive devices may not be as critical as previously assumed for AIML’s practical deployment.