LeafSplit is being benchmarked against established models on real low-data scientific tasks. Explore our experiments, methodology, and results.
The testing of Neural Network automatically built by Neural Network builder showed great results. The graphs with comporison classical neural network and automatically built with our architecture achieved impressive training stability and energy efficiency.
All models reach 100% test accuracy on the Wine benchmark — Boolear v8 gets there with just 16 parameters, 2.6× fewer than a Linear Softmax baseline.
Boolear v8 reaches 91% of Linear Softmax's accuracy on MNIST using just 1.8% of its parameters — a 56× compression, with a look at where the trade-off starts to bite.