Spiking Neural Networks
summary
Survey of training methods, measured tradeoffs and working tools: surrogate‑gradient backprop appears viable and, when combined with Batch Normalization Through Time (BNTT), cuts inference to 25–50 timesteps (from thousands in conversion) while achieving image‑level accuracy close to ANNs and estimated ~9× energy savings. A Spike Activation Map (SAM) visualizer uses short inter‑spike intervals to highlight discriminative features. Results include video/segmentation demos on DVS inputs, hybrid conversion+fine‑tuning for ImageNet, and improved robustness to noise and adversarial perturbations.
AI-written summary · about
chapters (13)
- 0:00Introduction to SNNs
- 1:37Evolution of Neural Nets
- 6:59Need for Efficient AI
- 12:10Brain Inspired Computing
- 16:07ANN vs SNN Mechanisms
- 22:37SNN Coding Techniques
- 27:47STDP Learning Rules
- 46:53ANN to SNN Conversion
- 1:05:16Backpropagation in SNNs
- 1:24:00Batch Norm Through Time
- 1:35:08Segmentation and SNNs
- 1:41:40SNN Interpretability
- 1:50:46Robustness and Conclusion
related talks
29mDeep Learning, Neuroscience and the future of AI
47mInterpreting Deep Neural Networks (DNNs)
55mNetworks of Spiking Neurons Learn to Learn and Remember
15mSMT-Based Modeling and Verification of Spiking Neural Networks: A Case Study
31mLarge Scale Stochastic Training of Neural Networks
13mEnhanced EEG Data Classification with 3D SNN and On-Line Learning ESN
56mChannel Gating for Cheaper and More Accurate Neural Nets
51mTips & Tricks for Fast Neural Net Inference in Production