In an arXiv paper, "Evolving Machine Learning Algorithms From Scratch," they introduced the program known as AutoML-Zero, which has the potential to develop AI algorithms using low-level math concepts (and with little to no human interaction involved). Similar to the evolutionary process, it improves code in each cycle by comparing the performance of candidate algorithms it generated against hand-created algorithms. During the process, the program edits, deletes, or replaces some of the code of the top-performing algorithms, which get added back into the population. The goal, according to computer scientists, is to scale up the work and forge new ML concepts that even researchers cannot find.
This story first appeared in Inside AI.