Take the 2-minute tour ×
Data Science Stack Exchange is a question and answer site for Data science professionals, Machine Learning specialists, and those interested in learning more about the field. It's 100% free, no registration required.

When I started with artificial neural networks (NN) I thought I'd have to fight overfitting as the main problem. But in practice I can't even get my NN to pass the 20% error rate barrier. I can't even beat my score on random forest!

I'm seeking some very general or not so general advice on what should one do to make a NN start capturing trends in data.

For implementing NN I use Theano Stacked Auto Encoder with the code from tutorial that works great (less than 5% error rate) for classifying the MNIST dataset. It is a multilayer perceptron, with softmax layer on top with each hidden later being pre-trained as autoencoder (fully described at tutorial, chapter 8). There are ~50 input features and ~10 output classes. The NN has sigmoid neurons and all data are normalized to [0,1]. I tried lots of different configurations: number of hidden layers and neurons in them (100->100->100, 60->60->60, 60->30->15, etc.), different learning and pre-train rates, etc.

And the best thing I can get is a 20% error rate on the validation set and a 40% error rate on the test set.

On the other hand, when I try to use Random Forest (from scikit-learn) I easily get a 12% error rate on the validation set and 25%(!) on the test set.

How can it be that my deep NN with pre-training behaves so badly? What should I try?

share|improve this question

1 Answer 1

up vote 12 down vote accepted

The problem with deep networks is that they have lots of hyperparameters to tune and very small solution space. Thus, finding good ones is more like an art rather than engineering task. I would start with working example from tutorial and play around with its parameters to see how results change - this gives a good intuition (though not formal explanation) about dependencies between parameters and results (both - final and intermediate).

Also I found following papers very useful:

They both describe RBMs, but contain some insights on deep networks in general. For example, one of key points is that networks need to be debugged layer-wise - if previous layer doesn't provide good representation of features, further layers have almost no chance to fix it.

share|improve this answer
    
Thank you! Your answer really helped me! –  izhak Jul 17 at 8:43

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.