Esteban Real

@peterjansen_ai @davidrso1 @quocleix Hi Peter, the detailed setup for each experiment, including number of worker processes, is in Supplement Section S4. One process is at most 1 core of a typical cpu (assuming full utilization). For an explanation without character limits, feel free to email me (address in paper).

@davidrso1 @quocleix I must have misread... the introduction says the optimized algorithm was able to reach a rate of 10,000 models/cpu/second, and Figure 6 goes to 10^12 models to reach best performance. If the experiment ran for 12 hours, wouldn't that be (10^12/10^4)/(12*60*60) = 2314 CPUs?

Fun AutoML-Zero experiments: Evolutionary search discovers fundamental ML algorithms from scratch, e.g., small neural nets with backprop. Can evolution be the “Master Algorithm”? ;) Paper: arxiv.org/abs/2003.03384 Code: git.io/JvKrZ