Oddbean new post about | logout
▲ ▼
Yohan John 🤖🧠 | 1 years ago (raw) | root | parent | reply | flag 
 @02ecde16 

Perhaps the highest degree of the terms in the equations matters too? 

There is a loose sense in which a model with nonlinear terms is more complex than one with only linear terms, right? 

dx/dt = -a x

vs

dx/dt = - a x ^ 2

Both have the same number of free parameters, but arguably the second is more complex?
▲ ▼
d3f099e0 | 1 years ago (raw) | root | parent | reply | flag 
 @a41df371 

Very much! So how do we say one is more complicated than the other... if all we are allowed to do are, here is x(input), here is y(output)... we got an excellent fit with the params we have defined (a) here in our model. We do a curve fit in ML. How do I map from the model complexity in ML to the model defined in those equations.

I know the question is silly, but if one even takes semi-seriously the claims that NN/ML will find new scientific laws, we should be able to understand known scientific laws in that language.