Show HN: Connectome-inspired spiking network you can try in the browser (awareliquid-sparse-snn-demo.static.hf.space)

A new browser-based application showcases a connectome-inspired spiking neural network that enables users to draw digits for classification. Utilizing a sparse connection density of just 5% with leaky integrate-and-fire (LIF) neurons, the network processes inputs with only event-driven additions, significantly reducing computational demands. This model runs on ONNX Runtime Web, allowing for interactive experimentation without requiring heavy installations. The significance of this development lies in its potential for more energy-efficient AI systems. The model's architecture offers an impressive estimated energy savings of approximately 112 times compared to traditional dense feedforward networks, with spiking neurons using only 0.9 pJ per addition compared to 3.7 pJ for multiply-accumulate operations. However, it's essential to note that while the approach is innovative, it's classified as a research artifact and not yet ready for production use, with performance gaps of about 1.5 points on MNIST and 0.5 points on Fashion datasets. Ultimately, this project highlights the promise of leveraging biological insights for computational efficiency in machine learning applications.