How might we extract insight into the nature of the motion of the planets from this model?
Have it predict the location of the planets based on where you are (and what time it is).
and a natural language description of what’s really going on behind the scenes
Predict how far away these planets are.
And herein lies the core of the problem: in this hypothesized world we do not know the true laws of Newtonian mechanics, so we cannot generate a reward signal by comparing the output of our model to ground truth during training.
Seems like not being able to connect that NL description to what’s going on in the NN is a bigger deal.
Also see [1]
To do this, we can set up an evaluation procedure that uses a data set of hand-labelled pictures of cats and dogs, and then use machine learning to search for a policy that correctly labels them. In contrast we do not at present know how to design an algorithm from first principles that does the same thing.
All these systems were engineered via first-principles design
Can we taboo “first principles”? (And maybe design.)
we should not for a moment think of it as the only game in town.
Arguably, the thought “experiment” early in the post can be thought of as backwards. Is it that
‘search produces great results, but we don’t know/understand how things ’actually work″, or that
‘Search comes from saying ‘we don’t care how it works, just make something that works’
Have it predict the location of the planets based on where you are (and what time it is).
Predict how far away these planets are.
Seems like not being able to connect that NL description to what’s going on in the NN is a bigger deal.
Also see [1]
Can we taboo “first principles”? (And maybe design.)
Arguably, the thought “experiment” early in the post can be thought of as backwards. Is it that
‘search produces great results, but we don’t know/understand how things ’actually work″, or that
‘Search comes from saying ‘we don’t care how it works, just make something that works’
[1] This article seems relevant: https://www.quantamagazine.org/how-artificial-intelligence-is-changing-science-20190311/, though it seems like a case of ‘if you don’t do X* one way, you’ll do it another’.
*Modeling, perhaps. ‘Science seems to pay little attention to the hypothesis generating procedure.’ Or:
1. Find a correlation.
2. Check if there’s causation.
3. Repeat one and two for Better Models.