We recently benchmarked J4 Stochastic Learning (J4 SL) versus tri-layered neural networks and hyperparameter optimized neural networks, and documented the test methods and results in a whitepaper that we just published on our J4 AI website. The summary: J4 Stochastic Learning is about four times more accurate, two to three orders of magnitude faster, and uses two to three orders of magnitude less power. We believe these benefits extend to more complex training scenarios, such as the popular LLM’s built using the GPT approach. We also see the same improvements for inference, not just for model training. All these factors mean that J4 Stochastic Learning can deliver better model performance at a small fraction of the heavyweight costs required by the current LLM approach.
You can download the whitepaper here, and you can find more info at j4cap.com.
