An Event-Driven Classifier for Spiking Neural Networks Fed with Synthetic or Dynamic Vision Sensor Data

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Computer Vision

Background:

• Spiking Neural Networks (SNNs) offer energy-efficient, event-driven processing.
• Training SNN classifiers, especially with real-world Dynamic Vision Sensor (DVS) data, remains challenging.
• Existing methods may not fully leverage the temporal dynamics of spiking activity.

Purpose of the Study:

• To introduce a novel supervised methodology for training event-driven SNN classifiers.
• To enable SNN classifiers to effectively process data from DVS chips and synthetic inputs.
• To demonstrate the adaptability and performance improvements of the proposed classifier.

Main Methods:

• A supervised method using spiking activity from prior SNN layers to build histograms.
• Training the classifier in the frame domain via stochastic gradient descent.
• Compatibility with leaky integrate-and-fire neuron models for realistic SNN applications.

Main Results:

• Achieved highest reported classification accuracy on N-MNIST (97.77%) and Poker-DVS (100%) datasets.
• Improved performance of a previously reported SNN by 2% through retraining the output layer.
• Demonstrated effectiveness across synthetic and real DVS datasets (MNIST, N-MNIST, MNIST-DVS, Poker-DVS).

Conclusions:

• The proposed method effectively trains SNN classifiers for DVS data, achieving state-of-the-art results.
• The classifier can enhance existing SNNs trained with unsupervised methods.
• The approach is suitable for real-world applications, considering neural dynamics and hardware implementation aspects.