AI Paper Atlas

Research map / Architectures and efficiency / Spiking networks

Practical Mechanism for Fault-Tolerant Spiking Neural Networks via Simple Input Control Based on Learnable Fragmentation

Hyun-Jong Lee, Jae-Han Lim

ICML 2026 · Architectures and efficiency › Spiking networks · approach: Training deep spiking networks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add learnable input fragmentation under stuck-at faults (ablation)VGG-7 · CIFAR-10, 30% SAFaccuracy (%): 52.65 -> 69.36Improved
add learnable division lines (LD) (ablation)fragmentation mechanism · CIFAR-100 ResNet-18, 30% SAFaccuracy (%): 2.98 -> 26.38Improved
add learnable number of fragments (LN) on top of LD (ablation)fragmentation mechanism · MNIST MLP, 50% SAFaccuracy (%): 21.86 -> 34.31Improved
add balance loss (BL) on top of LD+LN (ablation)fragmentation mechanism · MNIST MLP, 50% SAFaccuracy (%): 34.31 -> 56.81Improved
add entropy-based decoding on top of LD+LN+BL (ablation)fragmentation mechanism · MNIST MLP, 50% SAFaccuracy (%): 56.81 -> 93.52Improved
add input fragmentation at low fault ratio on simple MNIST (ablation)MLP · MNIST, 10% SAFaccuracy (%): 97.58 -> 96.34 (LD)Hurt
fixed 4-way input fragmentation under 50% SAFMLP / VGG-7 / ResNet-18 · MNIST, CIFAR-10, CIFAR-100accuracy (%): 10.82 -> 51.05 (MLP); 21.43 -> 37.29 (VGG-7); 1.12 -> 8.28 (ResNet-18)Improved
fragmentation reduces membrane potential deviation from threshold under SAFMLP · MNIST, 50% SAFE[|u_t - threshold|]: 80.18 -> 18.64Improved
fragmentation lowers probability of vanishing gradient under SAFMLP · MNIST, 50% SAFPr(|grad| < 0.0001): 0.99 -> 0.29Improved
compare proposed mechanism with lightweight approaches on inference timeVGG-7 · CIFAR-10, 30% SAF, softwareinference time (s): 1187.24 vs 1170.69 (threshold tuning)No clear effect
proposed mechanism vs best prior benchmark under SAF on large modelsResNet-34 · Tiny-ImageNet, 30% SAFaccuracy (%): 10.15 -> 16.4Improved
evaluate under random weight faults (RWFs) against ECOCfragmentation mechanism · MNIST/FMNIST/CIFARaccuracy (%)Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy