![[2602.21321] Dynamic Symmetric Point Tracking: Tackling Non-ideal Reference in Analog In-memory Training](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21321] Dynamic Symmetric Point Tracking: Tackling Non-ideal Reference in Analog In-memory Training
Summary
This article presents a novel approach to dynamic symmetric point tracking in analog in-memory computing, addressing the challenges posed by non-ideal device properties during model training.
Why It Matters
As machine learning models become increasingly complex, optimizing their training on energy-efficient hardware like analog in-memory computing is crucial. This research provides a theoretical framework and practical solutions to improve training accuracy, making it relevant for advancing AI technologies.
Key Takeaways
- Introduces a dynamic method for estimating symmetric points during model training.
- Addresses the limitations of existing calibration methods that assume known symmetric points.
- Demonstrates the effectiveness of the proposed method through numerical experiments.
Computer Science > Machine Learning arXiv:2602.21321 (cs) [Submitted on 24 Feb 2026] Title:Dynamic Symmetric Point Tracking: Tackling Non-ideal Reference in Analog In-memory Training Authors:Quan Xiao, Jindan Li, Zhaoxian Wu, Tayfun Gokmen, Tianyi Chen View a PDF of the paper titled Dynamic Symmetric Point Tracking: Tackling Non-ideal Reference in Analog In-memory Training, by Quan Xiao and 4 other authors View PDF HTML (experimental) Abstract: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 accuracy. 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. ...
