Oddbean new post about | logout
▲ ▼
jonny (good kind) | 1 years ago (raw) | root | parent | reply | flag 
 @02ecde16
Not VC dimension and related measures?
https://en.m.wikipedia.org/wiki/Vapnik%E2%80%93Chervonenkis_dimension
▲ ▼
7e50b099 | 1 years ago (raw) | root | parent | reply | flag 
 @ab7cf35b @02ecde16 

imho: Scientific models are a bit different than ML models -- they are closer to generator programs than to classifiers or function approximators. 

If theories are closer to programs, they are better defined by Kolmogorov complexity;  classifiers and function approximators are better defined (as mentioned) by VC dimension or Rademacher complexity... 

Theories also call each other recursively, like Maxwell's equations; another reason to consider them 'programs'.
▲ ▼
d3f099e0 | 1 years ago (raw) | root | parent | reply | flag 
 @5ff7031a @ab7cf35b 

I completely agree with this!

It was just an exercise to check my own sanity as well as see if others think in similar ways, vis-a-vis ML/NN systems will discover physical laws and all that nonsense.
▲ ▼
d3f099e0 | 1 years ago (raw) | root | parent | reply | flag 
 @ab7cf35b 

Sure... VC, Rademacher, whatever floats your boat. 

I give you a physics model, you give me its complexity. So far, I don't see anything like that meaningfully possible.