Efficient Calibration for RRAM-based In-Memory Computing using DoRA

2025-04-02

AI Summary

Conductance drift in RRAM-based in-memory computing severely degrades inference accuracy, yet existing retraining approaches are hindered by RRAM’s high write energy, latency, and limited endurance. Method: This paper proposes a lightweight calibration framework based on Decomposed Low-Rank Adaptation (DoRA), the first to adapt DoRA to RRAM in-memory computing: it stores only minimal calibration parameters (2.34% parameter update) in on-chip SRAM and compensates critical weights without any RRAM write operations. Contribution/Results: Leveraging weight sensitivity analysis and ultra-lightweight calibration—requiring only 10 samples—it restores 69.53% Top-1 accuracy on ResNet-50/ImageNet-1K. The method reduces calibration energy by >90% and shortens latency by two orders of magnitude, enabling efficient, rapid, and reliable zero-write calibration.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Adversarial Attacks & Robustness
• Intelligent Robots: Learning & Optimization for ROB

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Resistive In-Memory Computing (RIMC) offers ultra-efficient computation for edge AI but faces accuracy degradation due to RRAM conductance drift over time. Traditional retraining methods are limited by RRAM's high energy consumption, write latency, and endurance constraints. We propose a DoRA-based calibration framework that restores accuracy by compensating influential weights with minimal calibration parameters stored in SRAM, leaving RRAM weights untouched. This eliminates in-field RRAM writes, ensuring energy-efficient, fast, and reliable calibration. Experiments on RIMC-based ResNet50 (ImageNet-1K) demonstrate 69.53% accuracy restoration using just 10 calibration samples while updating only 2.34% of parameters.

Problem

• Addresses RRAM conductance drift in edge AI computing
• Reduces energy and latency in RRAM calibration
• Restores model accuracy with minimal parameter updates

Innovation

• DoRA-based calibration framework compensates influential weights
• Minimal calibration parameters stored in SRAM
• Eliminates in-field RRAM writes for efficiency