Research map / Architectures and efficiency / Spiking networks
Difference Predictive Coding for Training Spiking Neural Networks
Ville Karlsson, Nicklas Fianda, Joni-Kristian Kämäräinen
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 changed | Where | Measured | Effect |
|---|---|---|---|
| increase scheduler cycle length n from 3 to 7 (a=1.0) (ablation) | DiffPC · Numerical precision vs PC (128-200-10, random data) | mean absolute difference: 0.1506 -> 0.0117 | Improved |
| decrease decay value a from 1.0 to 0.1 at n=3 (ablation) | DiffPC · Numerical precision vs PC (128-200-10, random data) | mean absolute difference: 0.1506 -> 0.0168 | Improved |
| decrease decay value a from 1.0 to 0.1 at n=7 (ablation) | DiffPC · Numerical precision vs PC (128-200-10, random data) | mean absolute difference: 0.0117 -> 0.0072 | Improved |
| use spike-based DiffPC instead of backpropagation (long variant) | CNN 10C5S2-5C5S2-50FC-30FC-10FC · CIFAR-10 | accuracy (%): 63.5 -> 65.6 | Improved |
| shorten cycle length from n=16 (DiffPC-Long) to n=12 (DiffPC-Efficient) (ablation) | DiffPC on CIFAR-10 CNN · CIFAR-10 | accuracy (%): 65.6 -> 63.3 | Hurt |
| shorten cycle length from n=16 to n=12 (ablation) | DiffPC on CIFAR-10 CNN · CIFAR-10 (error propagation) | bits per neuron: 1.9 -> 0.7 | Improved |
| replace floating-point PC messages with ternary spikes (DiffPC-L vs PC-SE) | predictive coding on MNIST MLP · MNIST (error propagation) | bits per neuron: 960 -> 0.18 | Improved |
| use ternary-spike DiffPC instead of standard PC (smaller model) | predictive coding on MNIST MLP · MNIST (error propagation) | bits per neuron: 960 -> 0.08 | Improved |
| DiffPC-L instead of standard PC-SE | predictive coding on MNIST MLP · MNIST | accuracy (%): 98.3 -> 99.3 | Improved |
| DiffPC spike-based communication versus standard PC on timesteps | predictive coding on MNIST MLP · MNIST (error propagation) | timesteps: 15 -> 120 | Hurt |
| DiffPC with larger cycle length n on MNIST and Fashion-MNIST (ablation) | DiffPC · MNIST and Fashion-MNIST | accuracy (%) | Improved |
| reduce decay a below 1.0 at small n (ablation) | DiffPC · MNIST and Fashion-MNIST | accuracy (%) | Improved |
Closest papers by meaning
- Efficiently Training Time-to-First-Spike Spiking Neural Networks from Scratch (ICML 2026 · similarity 0.88)
- Direct Training for Spiking Neural Networks: Faster, Larger, Better (AAAI 2019 · similarity 0.87)
- SpikeGrad: An ANN-equivalent Computation Model for Implementing Backpropagation with Spikes (ICLR 2020 · similarity 0.87)
- Efficient Training of Deep Spiking Neural Networks with Input-Driven Derivative-Free Updates (NeurIPS 2026 · similarity 0.87)
- Advancing Training Efficiency of Deep Spiking Neural Networks through Rate-based Backpropagation (NeurIPS 2024 · similarity 0.86)
- Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training (NeurIPS 2025 · similarity 0.86)
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.
