Gradient-Free Continual Learning in Spiking Neural Networks via Inter-Spike Interval Regularization
New method uses spike timing regularity to prevent AI forgetting, enabling efficient learning on neuromorphic chips.
A research team led by Samrendra Roy and Kazuma Kobayashi has introduced ISI-CV, a breakthrough method for enabling Spiking Neural Networks (SNNs) to learn continuously without forgetting. The core innovation is a gradient-free synaptic importance metric derived from the regularity of a neuron's firing pattern—specifically, the Coefficient of Variation (CV) of its Inter-Spike Intervals. Neurons that fire with high regularity (low CV) are deemed to encode stable, task-relevant features and are protected from being overwritten during new learning. In contrast, neurons with irregular firing are allowed to adapt freely to new tasks. This approach sidesteps the need for backpropagation, making it natively compatible with neuromorphic hardware that uses only spike time counters and integer arithmetic.
In rigorous testing across four benchmarks, including Split-MNIST and real Dynamic Vision Sensor (DVS) event data, ISI-CV demonstrated exceptional performance. It achieved near-perfect scores with zero forgetting (AF = 0.000) on Split-MNIST and Split-FashionMNIST, and significantly outperformed gradient-based methods like Elastic Weight Consolidation on the challenging N-MNIST dataset with real neuromorphic data. On N-MNIST, where gradient-based methods failed, ISI-CV maintained high accuracy (AA = 0.820) with low forgetting (AF = 0.221). The method's hardware-friendly design and proven efficacy solve a critical roadblock in deploying adaptable, low-power AI systems in dynamic real-world environments, from industrial monitoring to edge devices.
- ISI-CV is the first gradient-free synaptic importance metric for SNN continual learning, using Inter-Spike Interval regularity (Coefficient of Variation) to protect critical neurons.
- It achieved near-zero forgetting (AF = 0.000) on Split-MNIST and the highest accuracy (AA = 0.820) on real neuromorphic DVS data (N-MNIST), outperforming gradient-based methods.
- The method requires only spike time counters and integer arithmetic, making it natively deployable on existing neuromorphic chips that lack backpropagation support.
Why It Matters
This unlocks efficient, lifelong learning for AI on low-power neuromorphic hardware, enabling smarter sensors and edge devices that adapt without forgetting.