Topology-Aware Recognition Framework Based on Ultra-High On/Off Ratio Memristors for Handwritten Digit Recognition

Authors:

Abstract
Conventional passive memristor arrays face crosstalk challenges in on-chip neuromorphic computing, while 1T1R/1S1R architectures suffer from limited scalability and area overhead. Here, we propose a topology-aware recognition framework for crosstalk-free on-chip compute-in-memory (CIM) using high-performance Pt/In2O3(Cr-doped)/W memristors. The devices demonstrate exceptional metrics: an ultra-high on/off ratio (>10⁸), retention exceeding 107 s, and endurance over 10⁵ cycles. Integrated into passive arrays, these memristors enable a novel paradigm that converts image pixel intensities into spatiotemporal pulse sequences encoding topological features (spatial brightness relationships). These features are mapped to 32 × 32 crossbar arrays, where recognition is determined by counting diagonal-activated devices—ensuring single-device activation per row/column to eliminate crosstalk. The system achieves 99.8% accuracy on MNIST and scalability to 6.59 PB (with 10% read margin), outperforming existing on-chip networks. This work bridges device innovation with algorithmic co-design, advancing energy-efficient neuromorphic hardware for pattern recognition.
Funding Information
  • Natural Science Foundation of Zhejiang Province (No. LDT23F0401)
  • National Natural Science Foundation of China (No.62204219)
  • open research fund of Suzhou Laboratory (No.SZLAB-1208-2024-TS012)

This publication has 24 references indexed in Scilit: