[sirtoffski]

Apologies in advance if my terminology is not entirely correct - no formal CS background.

I wonder what implications, if any, this method has on the way science is done.

Perhaps I’m out of date, but I have always thought that with machine learning, with have little insight into how the algorithm does exactly what it does. We know how to train it, how it works - in terms of expected output given a certain input. But is not knowing exactly how the result is achieved a problem?

For modelling airplane air resistance, etc - it may not matter much. For fundamental physics research, I am not so sure. If replicating and verifying experimental results is so important to the “scientific method” - how would one approach confirming the results obtained via ML? Does independently building another ML algorithm count?