Computational AI

A new optimization algorithm, B67, is particularly applicable to the neural-network training problem, and its fast performance results in orders of magnitude of savings in computational effort.

This algorithm is the result of a theoretical breakthrough after more than 35 years of work on this problem. This is neither a heuristic, nor a variant of existing algorithms.

AI-training Convergence on ResNet50 (click image for details)

See the rest of this site for explanation and examples including:

Notes on AI training,
Tensorflow Demos (ResNet50 on 1M images, transformer with >100M parameters, etc.)