What is the network size limit in PyBrain?

Question

I have recently constructed a neural network in PyBrain, but ran in to a problem of my network being to big.

• So what are the exact limits in PyBrain?

• How can I construct a network with 750000 inputs, 3 outputs and a working ammount of hidden nodes(currently it only works with 35-40 hidden nodes, which I feel would be a lot too little)?

• Are there any other ways to construct image recognition(in my case basic, but not identical(random solid colors and sizes) shape recognition - square, circle and an equilateral triangle) than using MLP with n input nodes where n = image.width * image.height * 3?

Answer 1

With that many nodes even if there would be no computation errors the amount of time which network would need to perform single training epoch would be enormous. I don't believe that there is specific limit to the network size in PyBrain other than common sense and your computer capabilities.

Also the idea of neural network is to simplify complex problems. So you have to shrink input data. And I mean shrink it a lot, for your purpose about 100 input nodes would be more then enough.

I recommend some source material for you about choosing input data, and network architecture:

• Localized Precipitation Forecasts from a Numerical Weather Prediction Model Using Artificial Neural Networks

• Great answer about network architecture

• General article on Wikipedia

It's a bit a matter of experience to chose appropriate input data, but you will get to it.

Hope that was helpful.