SCIB: Robust representation learning for spiking neural networks based on a conditional information bottleneck
Start here
The researchers propose SCIB, a robust learning objective for spiking neural networks (SNNs) based on a conditional information bottleneck, which replaces the compression term I(X;S) of the conventional variational information bottleneck with the conditional mutual information I(X;S|Y). In experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 against environmental noise, hardware noise, and adversarial attacks, SCIB is more robust than standard SNN and IB baselines under most conditions, improves adversarial accuracy without adversarial training, and reduces SynOps inference energy by 24.9%.
Generated by AI from reports · Updated 23 days ago
Attention and developments
Attention, hourly
Now 16, 0% in 24 hours·Peak 16 (Nov 6 03:00)
Related research
This event's paper
- Robust Representation Learning for Spiking Neural Networks based on Conditional Information Bottleneck
arXiv · 2026-10-20 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningSCIB: Robust Representation Learning for Spiking Neural Networks Based on the Conditional Information Bottleneck
The researchers propose SCIB, a robust learning objective for SNNs based on the conditional information bottleneck, replacing the compression term I(X;S) of the conventional variational information bottleneck with the conditional mutual information I(X;S|Y). In experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 against environmental noise, hardware noise, and adversarial attacks, SCIB is more robust than standard SNNs and IB baselines under most conditions, improving adversarial accuracy without adversarial training, and reducing SynOps inference energy by 24.9%.