Hello,
When using a Deep Belief network (DBN) for images classification, how could we choose the number of hidden layers (number of Restricted Boltzmann Machines) and nodes (RBM size) in a hidden layer? (For example in case of an input size of 19200 pixels )
I will be grateful for any help you can provide :)
Wafa
There are several papers on the number of hidden layers needed for universal approximation (e.g., Le Roux and Benjio, Montufar) of "narrow" DBNs.
However, you should take into account the amount of training data you have vs. the number of free parameters.
Y(J)S
I think the best way is cross-validation using supervised method with training data. It means that using trial and error and monitor the performance of your topology.
please see
http://www.cs.toronto.edu/~asamir/cifar/cifarss-marlin.pdf
http://arxiv.org/pdf/1206.5533v2.pdf,
http://cseweb.ucsd.edu/~dasgupta/254-deep/nakul.pdf
https://arxiv.org/pdf/1811.12587
I think you need to check this link which will help you
https://medium.com/swlh/what-are-rbms-deep-belief-networks-and-why-are-they-important-to-deep-learning-491c7de8937a
Dear Yaakov J Stein , Javad Sharifi , Mahmoud Khaled Abd-Ellah thank you so much.
No reason to thank me - Javad's answer was the best (although I am not a fan of trial and error methods).
Y(J)S
- Badges
- Science topic