The Need

Spiking Neural Networks (SNNs) promise ultra-low-power, low-latency AI for edge and neuromorphic computing, but their adoption is constrained by fundamental training challenges. SNN performance is highly sensitive to how training data are collected (e.g., lighting, sensor settings, noise), leading to poor generalization when conditions change. Collecting comprehensive neuromorphic datasets to address this is costly, slow, and often impractical, creating a critical barrier to reliable deployment of SNN-based systems.

The Technology

OSU engineers have developed a platform-agnostic, AI-driven approach for improving the training and robustness of spiking neural networks. It leverages generative AI to learn the underlying structure of existing neuromorphic datasets and produce high-quality synthetic spike data that systematically broadens training coverage. By intelligently augmenting and selecting training samples, the method reduces sensitivity to data-collection conditions and strengthens generalization, without requiring new sensors, changes to SNN architectures, or extensive additional data collection.

Commercial Applications

• Edge AI systems using neuromorphic processors (vision, sensing, autonomy)
• Event-based computer vision for robotics, drones, and autonomous vehicles
• Defense and aerospace sensing systems operating under variable conditions
• Low-power AI solutions for IoT and embedded intelligence platforms

Benefits/Advantages

• Improved robustness: Significantly reduces performance degradation across changing sensor and environmental conditions
• Data-efficient: Extracts more value from existing neuromorphic datasets, minimizing costly data collection
• Hardware-agnostic: Compatible with current SNN training frameworks and neuromorphic processors
• Low overhead: Enhances model performance without materially increasing training or inference costs