Research map / Label-efficient and robust learning / Recognition and long-tail
Neuronal Competition Groups with Supervised STDP for Spike-Based Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Neuronal Competition Groups (NCGs), intra-class WTA groups with two-compartment thresholds for competition regulation, added to spiking classification layers trained with supervised STDP (SSTDP, S2-STDP) on top of Hebbian feature extractors.
Problem
Supervised STDP classification layers with first-spike coding struggle with unbalanced winner-takes-all competition when multiple neurons per class are used, limiting the learning of varied class-specific patterns.
Approach
Group neurons per class into an NCG with intra-class lateral inhibition (no inter-class WTA), and regulate competition with a fixed test threshold for inference plus an adaptive training threshold that is only active on samples of the neuron's own class.
Key results
- S2-STDP+NCG reaches 99.17% on MNIST, 91.86% on Fashion-MNIST and 79.55% on CIFAR-10 with SoftHebb-CNN features, outperforming the S2-STDP baseline (98.81%, 90.61%, 76.90%).
- With STDP-CSNN features, S2-STDP+NCG improves over S2-STDP by 5.33 pp on CIFAR-10 and 6.51 pp on CIFAR-100 (to 35.90%).
- Ablation on CIFAR-10 with STDP-CSNN shows two-compartment competition regulation (CR-2) reaches 65.51% vs 62.61% for M-5 without regulation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add NCG to S2-STDP | S2-STDP classification layer · MNIST, STDP-CSNN | accuracy: 97.74 ± 0.06 -> 98.92 ± 0.07 | Helped |
| Add NCG to S2-STDP | S2-STDP classification layer · Fashion-MNIST, STDP-CSNN | accuracy: 85.89 ± 0.27 -> 88.72 ± 0.23 | Helped |
| Add NCG to S2-STDP | S2-STDP classification layer · CIFAR-10, STDP-CSNN | accuracy: 61.08 ± 0.17 -> 66.41 ± 0.17 | Helped |
| Add NCG to S2-STDP | S2-STDP classification layer · CIFAR-100, STDP-CSNN | accuracy: 29.39 ± 0.19 -> 35.90 ± 0.42 | Helped |
| Use five neurons per class instead of one (ablation) | S2-STDP+NCG without competition regulation or neuron labeling · Fashion-MNIST and CIFAR-10, both feature extractors | accuracy: 85.89 ± 0.27 -> 86.74 ± 0.25; 90.61 ± 0.19 -> 91.25 ± 0.20; 61.08 ± 0.17 -> 62.61 ± 0.27; 76.90 ± 0.27 -> 78.29 ± 0.27 | Helped |
| Add two-compartment threshold competition regulation (ablation) | S2-STDP+NCG with M=5 · Fashion-MNIST and CIFAR-10, both feature extractors | accuracy: 86.74 ± 0.25 -> 87.76 ± 0.16; 91.25 ± 0.20 -> 91.33 ± 0.22; 62.61 ± 0.27 -> 65.51 ± 0.26; 78.29 ± 0.27 -> 78.78 ± 0.15 | Helped |
| Replace two-compartment regulation with one threshold (ablation) | S2-STDP+NCG with M=5 · Fashion-MNIST and CIFAR-10, both feature extractors | accuracy: 86.74 ± 0.25 -> 86.77 ± 0.22; 91.25 ± 0.20 -> 85.16 ± 5.34; 62.61 ± 0.27 -> 64.73 ± 0.41; 78.29 ± 0.27 -> 77.35 ± 0.22 | Mixed |
| Add neuron labeling to the two-compartment threshold variant (ablation) | S2-STDP+NCG with M=5 · Fashion-MNIST and CIFAR-10, both feature extractors | accuracy: 87.76 ± 0.16 -> 88.72 ± 0.23; 91.33 ± 0.22 -> 91.86 ± 0.14; 65.51 ± 0.26 -> 66.41 ± 0.17; 78.78 ± 0.15 -> 79.55 ± 0.23 | Helped |
| Add neuron labeling to the one-threshold variant (ablation) | S2-STDP+NCG with M=5 and CR-1 · Fashion-MNIST and CIFAR-10, both feature extractors | accuracy: 86.77 ± 0.22 -> 87.14 ± 0.41; 85.16 ± 5.34 -> 89.24 ± 0.89; 64.73 ± 0.41 -> 65.46 ± 0.40; 77.35 ± 0.22 -> 78.67 ± 0.19 | Mixed |
| Replace two-compartment threshold regulation with dropout (ablation) | S2-STDP+NCG with M=5 and neuron labeling · Fashion-MNIST and CIFAR-10, both feature extractors | accuracy: 88.72 ± 0.23 vs 87.33 ± 0.20; 91.86 ± 0.14 vs 91.34 ± 0.08; 66.41 ± 0.17 vs 63.15 ± 0.11; 79.55 ± 0.23 vs 77.98 ± 0.21 | Hurt |
| Use S2-STDP instead of SSTDP with NCG | NCG classification layer · All four datasets and both feature extractors | accuracy | Helped |
Surprising findings
- Learning thresholds with a single threshold per neuron (CR-1) is not mandatory for learning various patterns; fixed test thresholds for decision-making with threshold adaptation only for regulation work better.
- R-STDP's advantage over S2-STDP on hard features is closed by NCG, which uses four times fewer neurons per class.
Limitations
- Computational, parameter and hardware overhead scales linearly with neurons per class.
- More neurons per class increases specialization on the training set and limits generalization, especially with richer input features (SoftHebb-CNN).
- NCG applies only to the output classification layer; no multi-layer WTA extension is shown.
Experimental setup
Datasets: MNIST Fashion-MNIST CIFAR-10 CIFAR-100
Benchmarks: MNIST Fashion-MNIST CIFAR-10 CIFAR-100
Metrics: accuracy (mean ± std over 10-fold cross-validation)
Baselines: R-STDP SSTDP S2-STDP STiDi-BP EMSTDP Sym-STDP S4NN
Builds on: S2-STDP SSTDP R-STDP STDP-CSNN SoftHebb
Closest papers by meaning
- A Plasticity-Centric Approach to Train the Non-Differential Spiking Neural Networks (AAAI 2018 · similarity 0.86)
- New Efficient Multi-Spike Learning for Fast Processing and Robust Learning (AAAI 2020 · similarity 0.86)
- Heterogeneous Neuronal and Synaptic Dynamics for Spike-Efficient Unsupervised Learning: Theory and Design Principles (ICLR 2023 · similarity 0.86)
- Spike-Train Level Backpropagation for Training Deep Recurrent Spiking Neural Networks (NeurIPS 2019 · similarity 0.86)
- Enabling Deep Spiking Neural Networks with Hybrid Conversion and Spike Timing Dependent Backpropagation (ICLR 2020 · similarity 0.86)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.84)
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.