I would like to create an ANN to solve some AI problems. How many hidden layers do I have to put? I have read that one hidden layer is sufficient to solve any problem... cheers

theoretically yes, but only if you train it properly, avoiding local minima during the training process.

Quote: Original post by sidhantdash
theoretically yes, but only if you train it properly, avoiding local minima during the training process.

could you be a lot more precise, because i did not get what you said

What I meant was that one hidden layer is enough to solve any problem, but only theoretically. Most often, during the training process (using back-propagation, or some similiar gradient descent algorithm) the network gets stuck in a local minima. So you then end up with sub-optimal weights, thereby failing to solve your problem. So you need to train your net properly, to ensure that it doesnt get stuck in a local minima. It doesnt mean you have to increase the number of hidden layers, it just means you need to look at the training process again. You also will need to pre-process your data, for best results.

One might use methods like genetic algorithms, simulated annealing etc to break out of local minima.

Hope that helps.

Quote: Original post by sidhantdash
What I meant was that one hidden layer is enough to solve any problem, but only theoretically. Most often, during the training process (using back-propagation, or some similiar gradient descent algorithm) the network gets stuck in a local minima. So you then end up with sub-optimal weights, thereby failing to solve your problem. So you need to train your net properly, to ensure that it doesnt get stuck in a local minima. It doesnt mean you have to increase the number of hidden layers, it just means you need to look at the training process again. You also will need to pre-process your data, for best results.

One might use methods like genetic algorithms, simulated annealing etc to break out of local minima.

Hope that helps.

ok, i get it