Dynamic Symmetric Point Tracking: Tackling Non-ideal Reference in Analog In-memory Training

Analog in-memory computing (AIMC) performs computation directly within resistive crossbar arrays, offering an energy-efficient platform to scale large vision and language models. However, non-ideal analog device properties make the training on AIMC devices challenging. In particular, its update asymmetry can induce a systematic drift of weight updates towards a device-specific symmetric point (SP), which typically does not align with the optimum of the training objective. To mitigate this bias, most existing works assume the SP is known and pre-calibrate it to zero before training by setting the reference point as the SP. Nevertheless, calibrating AIMC devices requires costly pulse updates, and residual calibration error can directly degrade training performance. In this work, we present the first theoretical characterization of the pulse complexity of SP calibration and the resulting estimation error. We further propose a dynamic SP estimation method that tracks the SP during model training, and establishes its convergence guarantees. In addition, we develop an enhanced variant based on chopping and filtering techniques from digital signal processing. Numerical experiments demonstrate both the efficiency and effectiveness of the proposed method.

Analog in-memory computing (AIMC) is a promising hardware platform for reducing the energy cost of training large neural networks, since it performs matrix-vector multiplications directly in resistive crossbar arrays. However, analog pulse updates are not ideal: an up pulse and a down pulse can change a weight by different magnitude. This update asymmetry creates a systematic weight drift toward a device-dependent symmetric point (SP), which is usually different from the optimal model weights that minimize the training loss. Existing analog training methods often assume that this SP is known and shifted to zero before training, but this zero-shifting calibration consumes many pulse updates, and any remaining estimation error can degrade training performance. This paper first quantifies the trade-off between pulse budget and SP estimation accuracy, showing why static calibration becomes costly for high-precision devices. We then propose RIDER, a dynamic method that tracks the SP during model training, and establish its convergence guarantees. We further introduce E-RIDER, which uses chopping and filtering to improve SP tracking performance and reduce practical overhead. Experiments show that E-RIDER is robust to nonzero SPs and improves analog training accuracy with fewer overall pulse updates.