Relation between input data points and hyper parameters that needs to be tuned

2 visualizaciones (últimos 30 días)
Hi All,
Can anyone please let me know the relationship between the number of input data points and the hyperparameters/number of layers that needs to be present in any machine learning model?
Thanks for your time and help

Respuesta aceptada

Greg Heath
Greg Heath el 9 de Ag. de 2018
Editada: Greg Heath el 9 de Ag. de 2018
[ I N] = size(input)
[ O N ] = size(target)
% (MATLAB DEFAULT)
Ntst = round(0.15*N)
Nval = Ntst
Ntrn = N-(Ntst+Nval)% ~ 0.7*N
% Design parameters
Ndes = Ntrn*O % No. of design equations ~ 0.7*N*O
H % No. of hidden nodes for I-H-O net
Nw = (I+1)*H+(H+1)*O % No. of unknown weights
Require Ndes >= Nw ==> H <= Hub = (Ntrn*O-O)/(I+O+1)
Desire Ndes >> Nw ==> H << Hub
My typical goal: Minimize H subject to the requirement
MSE < = 0.01*var(target',1) % Rsquare >= 0.99
My approach:
1. Apply the requirement to the training data
2. Loop over H to find the minimum H to satisfy the
requirement.
I have hundreds of examples in the NEWSGROUP comp.soft-sys.matlab as well as ANSWERS.
Hope this helps
Thank you for formally accepting my answer
Greg
  1 comentario
Venkat
Venkat el 9 de Ag. de 2018
Hi Greg,
Thanks for your time and input. I understood what you have said. My problem is I am using CNN. My inputs are images of size 16x512 and I have 30,000 image samples per class, totaling 60,000 representing my 2 classes.
In order to decide on the number of layers in CNN along with the number of filters in each convolutional layer, I am doing an iterative process, but that is going to take a long time. So I am trying to see if I can derive some generic numbers I can start with rather than iterating from 1 to N as far as the number of filters is concerned.
Can I apply the same rule? If yes, can you please explain a little bit more
Thanks

Iniciar sesión para comentar.

Más respuestas (1)

Greg Heath
Greg Heath el 11 de Ag. de 2018
Each case is different. However, things tend to be relatively straightforward if you have at least as many training equations as you have unknowns.

Categorías

Más información sobre Image Data Workflows en Help Center y File Exchange.

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by