[Paper] Effects of Introducing Synaptic Scaling on Spiking Neural Network Learning

Source: arXiv - 2601.11261v1

Overview

The paper investigates how adding synaptic scaling—a biologically‑inspired homeostatic mechanism—to spiking neural networks (SNNs) affects their ability to learn visual patterns. By combining spike‑time‑dependent plasticity (STDP) with different scaling rules in a winner‑take‑all (WTA) architecture, the authors show that an L2‑norm‑based scaling scheme can substantially boost classification accuracy on MNIST and Fashion‑MNIST, even after a single training epoch.

Key Contributions

• Hybrid plasticity framework: Integrated STDP with three variants of synaptic scaling (L1‑norm, L2‑norm, and max‑norm) inside a WTA SNN.
• Systematic ablation study: Explored the impact of neuron count, STDP time constant, and scaling method on performance.
• Empirical benchmark: Achieved 88.84 % accuracy on MNIST and 68.01 % on Fashion‑MNIST using only 400 excitatory + 400 inhibitory neurons and one epoch of training.
• Open‑source implementation: Provided a Python codebase that reproduces the experiments, facilitating reproducibility and further research.

Methodology

1. Network architecture – A two‑layer WTA circuit:

   • Excitatory layer (E) receives pixel spikes encoded via Poisson firing.
   • Inhibitory layer (I) implements lateral competition so that only a subset of excitatory neurons fire for a given input.

2. Learning rules –

   • STDP updates each synapse based on the relative timing of pre‑ and post‑spikes, controlled by a time constant τ_STDP.
   • Synaptic scaling periodically renormalizes all incoming weights of a neuron to keep overall synaptic strength bounded. Three scaling strategies were tested:
     • L1‑norm scaling (sum of absolute weights = 1)
     • L2‑norm scaling (Euclidean norm = 1) – the best performer
     • Max‑norm scaling (largest weight = 1)

3. Training protocol –

   • Input images from MNIST/Fashion‑MNIST are converted to spike trains (rate‑coded).
   • The network processes each sample for a fixed simulation window; STDP and scaling are applied online.
   • After one epoch, the firing patterns of excitatory neurons are used as feature vectors for a simple linear classifier (softmax) to compute final accuracy.

4. Experimental variables – Number of neurons per layer (200–600), τ_STDP (10–30 ms), and scaling method.

Results & Findings

• L2‑norm scaling consistently outperformed L1 and max‑norm, suggesting that preserving the Euclidean magnitude of weight vectors better maintains the balance between excitation and inhibition.
• Increasing the neuron count beyond 400 gave diminishing returns, indicating that the WTA competition already extracts a compact, discriminative code.
• Shorter STDP time constants (more precise timing) slightly improved MNIST performance but hurt Fashion‑MNIST, highlighting dataset‑dependent optimal temporal windows.

Practical Implications

• Energy‑efficient AI on neuromorphic hardware: SNNs with homeostatic scaling can achieve respectable accuracy after a single pass, making them attractive for low‑power edge devices where training resources are scarce.
• Online, continual learning: Because scaling is applied continuously, the network can adapt to non‑stationary data streams without catastrophic forgetting—a key requirement for robotics and IoT sensors.
• Simplified training pipelines: Developers can replace costly back‑propagation with biologically plausible rules, reducing the need for large GPU clusters while still obtaining competitive results on classic vision benchmarks.
• Guidance for neuromorphic chip designers: The L2‑norm scaling rule is computationally cheap (just a vector norm and division) and can be implemented directly in analog/digital synapse circuits, informing future chip architectures.

Limitations & Future Work

• One‑epoch training: While impressive for a single pass, the reported accuracies still lag behind state‑of‑the‑art deep CNNs; extending to multi‑epoch or hybrid training could close the gap.
• Dataset scope: Only two relatively small, grayscale datasets were evaluated; performance on larger, color, or video data remains unknown.
• Scaling frequency: The study applied scaling at fixed intervals; adaptive scheduling based on network activity could further improve stability.
• Hardware validation: Experiments were conducted in software simulation; real‑world deployment on neuromorphic platforms (e.g., Loihi, SpiNNaker) is needed to confirm energy and latency benefits.

Authors

• Shinnosuke Touda
• Hirotsugu Okuno

Paper Information

• arXiv ID: 2601.11261v1
• Categories: cs.NE, cs.LG
• Published: January 16, 2026
• PDF: Download PDF