I aim at solving a classification problem using two-layered neural network with back propagation algorithm. Will my accuracy increase if I increase the no of neurons in the hidden layer of this two layered network? Is increase of neurons of any use in two layered neural network? Thanks in advance to all.
The model gets more adaptive, so it can learn smaller details. But this means not necessary, that your classifier is then better on the 'next dataset'. The model gets more affected to over-fitting and so the generalization of you classification model can also decrease, e.g. your classifier will work worse on the next dataset.
Dear Thomas and Paul,
Thanks for you answers. Both of your answers were pretty convincing. I just need a bit of clarification in deciding on the number of layer to go with and the neurons in each layer. Is there any algorithm, or logic behind deciding or choosing the no of layers and neurons?
Hi Manas,
one conceptual model of ANNS is the separation of classes using convex hulls in multi-dimensional space. To put it crudely, if lots of hyperplanes are required to separate many hulls and then recombine them for a final classification, you are likely to need more neurons in the hidden layer. You will be able to find some rules of thumb in the literature which may or may not help. They may have some mathematical justification which may or may not be sound. The best thing to do is to plan a good systematic series of experiments to explore the performance of different architectural combinations and use measures of statistical significance to see whether one is better than the others (Weka has a measure built in).
Sorry, Paul has illustrated the need for clarification of my previous post.
I was taking it for granted that the training set would not be used for evaluation.
Also, I was assuming that it would be understood that the statistical tests I was referring to would be used to determine whether the test set results were significantly different from each other for the different network architectures.
- Badges
- Science topic