Let us assume we have a d1-by-d matrix (lets us call this one A) for the input and a d2-by-d (B) matrix for the target of a neural network.
I see that the canonical correlation between the columns of A and B is all greater than 0 and less than 1 (in my particular example it is [0.8,0.5,0.1]) using Matlab's canoncorr(A,B) function.
Now I noticed that Matlab's "neural network fitting" application converges on a mean squared error of 160 using a variety of different algorithms and hidden layers.
Can this convergence limit be estimated using the cannoncorr function?
Dear Tristan,
Yes, most likely and to a certain extent, depending on how linearly correlated input and target would be, and depending on your algorithm implementing a CCA in order to bridging the input and target.
Your suggested estimation method seems to be a good start to approach reaching a further stable training.
Best regards,
J.
- Badges
- Science topic