| 講演抄録/キーワード |
| 講演名 |
2026-08-07 16:10
Quantization-Aware Fault-Injected Level Adjustment for Reliable Neural Network Deployment on Memristor Crossbars ○Mizanur Rahman・Md. Sihabul Islam・Taisho Sasada・Michiko Inoue(NAIST) CPSY2026-33 DC2026-33 RECONF2026-33 |
| 抄録 |
(和) |
Memristor crossbar arrays offer a compelling path to energy-efficient neural network acceleration, but permanent stuck-at faults (SAFs), conductance quantization errors, and stochastic device noise collectively threaten inference reliability. Existing methods rely on per-chip retraining or redundant hardware, leaving the joint effect of SAFs, conductance quantization, and write noise unaddressed. We propose FILA-QAT (Fault-Injected Level-Adjusted Quantization-Aware Training), a unified framework that enforces fault constraints directly in the integer level domain to produce inherently fault-tolerant, quantization-efficient, and noise-resilient inference. During deployment, FILA-QAT applies level adjustment (LA) within the mapping phase; noise robustness is evaluated by perturbing the programmed conductances with log-normal write noise, without per-device retraining or redundant circuitry. Benchmarked against PTQ (post-training quantization) across multiple architectures at fault rates from 5% to 30%, FILA-QAT consistently outperforms all hardware-realizable (PTQ-based) baselines, achieving up to 3.21% accuracy gain at a 30% fault rate while showing no measurable write-noise degradation at σ= 0.05, reducing the maximum average accuracy drop from 0.062% to 0.00%. |
| (英) |
Memristor crossbar arrays offer a compelling path to energy-efficient neural network acceleration, but permanent stuck-at faults (SAFs), conductance quantization errors, and stochastic device noise collectively threaten inference reliability. Existing methods rely on per-chip retraining or redundant hardware, leaving the joint effect of SAFs, conductance quantization, and write noise unaddressed. We propose FILA-QAT (Fault-Injected Level-Adjusted Quantization-Aware Training), a unified framework that enforces fault constraints directly in the integer level domain to produce inherently fault-tolerant, quantization-efficient, and noise-resilient inference. During deployment, FILA-QAT applies level adjustment (LA) within the mapping phase; noise robustness is evaluated by perturbing the programmed conductances with log-normal write noise, without per-device retraining or redundant circuitry. Benchmarked against PTQ (post-training quantization) across multiple architectures at fault rates from 5% to 30%, FILA-QAT consistently outperforms all hardware-realizable (PTQ-based) baselines, achieving up to 3.21% accuracy gain at a 30% fault rate while showing no measurable write-noise degradation at σ= 0.05, reducing the maximum average accuracy drop from 0.062% to 0.00%. |
| キーワード |
(和) |
Memristor Crossbar / Stuck-at Faults (SAF) / Quantization-Aware Training (QAT) / FILA-QAT / WBM Mapping / Noise Robustness / / |
| (英) |
Memristor Crossbar / Stuck-at Faults (SAF) / Quantization-Aware Training (QAT) / FILA-QAT / WBM Mapping / Noise Robustness / / |
| 文献情報 |
信学技報, vol. 126, no. 141, DC2026-33, pp. 97-102, 2026年8月. |
| 資料番号 |
DC2026-33 |
| 発行日 |
2026-07-30 (CPSY, DC, RECONF) |
| ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
| PDFダウンロード |
CPSY2026-33 DC2026-33 RECONF2026-33 |
|