Google Engineers Mutate AI to Make Systems Evolve Faster Than We Can Code Them

> The system starts off with a selection of 100 algorithms made by randomly combining simple mathematical operations. A sophisticated trial-and-error process then identifies the best performers, which are retained - with some tweaks - for another round of trials. In other words, the neural network is mutating as it goes.

This approach sounds quite similar to this neuroevolution paper [0] but perhaps that this sort of works if you have a system that has a large amount of memory that can handle simultaneously evolving these neural networks. Thus, the hardware requirements would unsurprisingly be demanding; at least for Google they'd just use their own infrastructure and training it on their TPU hardware.

This is the only use-case that makes sense to apply neuroevolution to search for the optimal network, unless you're Google or on their GCP infrastructure that is.

[0] http://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf

Here is a paper they referenced in the article:

2003) AutoML-Zero: Evolving Machine Learning Algorithms From Scratch - https://arxiv.org/abs/2003.03384.

Although I could not finish reading all of it yet, I think here algorithms mutating is another dimension which can mutate, which then effects of course new topologies, weights, and other factors.