Any sufficiently complex system acts as a black box when it becomes easier to experiment with than to understand. Hence, black-box optimization has become increasingly important as systems become more complex. - Google Vizier: a service for black-box optimization Golovin et al., KDD'17
I would rather a high level algorithm description as an output -- which could definitely be fed into some sort of compiler that ultimately outputs executable code.
I feel like going straight to executable code isn't solving the problem GP was interested in, which I believe to be the problem of transferring knowledge from machine to engineer in much the way an engineer would transfer it to another engineer.
An algorithm that outputs code without any high level understanding or documentation is about as useful to me in a large project as an intern who can copy-paste from Stack Overflow and produce volumes of code with no documentation, in the long term.
I suspect many (most?) algorithms are sufficiently complex as to make this completely infeasible, but hopefully I'm wrong!
Perhaps MT (machine teaching?) is the next evolution of ML.
My enthusiasm in this instance is probably tempered by the fact that image resizing is on the simple end of things we're using ML for, I'd think.
It's a two dimensional grid of data points. That's it. I mean, that's certainly not trivial (look at all the algorithms we've come up with just in the last 10-20 years! imagine all the people-hours!) but it pales in complexity to, say, weather models or automated scanning of PET scans for tumors or something.
Image the output of any given image sizing algorithm can be quickly assessed by eye so that's a very convenient feedback loop. As opposed to say, using ML to come up with proposed oil drilling locations where testing out each proposed drilling spot is a very expensive proposition.
Perhaps we can cut out the middleman here. Maybe the answer is not for ML models to come up with human-understandable algorithms. Perhaps the answer is for them to produce optimized code that implements the algorithms they've discovered.Disclaimer, in case it's not blindingly obvious - I am not versed in ML at all.