# Neural Network Hidden Neuron Selection Strategy

I'm trying to determine what is the best number of hidden neurons for my MATLAB neural network. I was thinking to adopt the following strategy:

• Loop for some values of hidden neurons, e.g. 1 to 40;
• For each NN with a fixed number of hidden neurons, perform a certain number of training (e.g. 40, limiting the number of epoch for time reasons: I was thinking to doing this because the network seems to be hard to train, the MSE after some epochs is very high)
• Store the MSE obtained with all the nets with different number of hidden neurons
• Perform the previous procedure more than 1 time, e.g. 4, to take into account the initial random weight, and take the average of the MSEs
• Select and perform the "real" training on a NN with a number of hidden neurons such that the MSE previously calculated is minimized

The MSE that I'm referring is the validation MSE: my samples splitting in trainining, testing and validation to avoid overfitting is 70%, 15% and 15% respectively)

Other informations related to my problem are:
fitting problem
9 input neurons
2 output neurons
1630 samples

This strategy could be work? Is there any better criterion to adopt? Thank you

Edit: Test done, so the result suggest me to adopt 12 neurons? (low validation MSE and number of neurons lower than 2*numberOfInputNeurons? but also 18 could be good...

• From the shape of the curve, I want to say to run with 6 or 7 since you 'should' be looking for the 'shoulder'. However, your validation and testing scores look to be doing surprisingly well even at the highest number of hidden nodes. Any chance you could try the test for even more nodes (upwards of 35 or 40) to see if you can start breaking your validation and testing sets? Jan 5 '15 at 22:27

A rule of thumb approach is:

• start with a number of hidden neurons equal (or little higher) that the number of features.
• In your case it would be 9. My suggestion is to start with 9*2 = 18 to cover a wider range of possibilities.
• Be sure your test and validation sets are selected "fairly": a random selection and varying the seed some number of times to test different configurations would be ok.

In general, a number of neurons equal to the number of features will tend to make each hidden neuron try to learn that special thing that each feature is adding, so will could say it is "learning each feature" separately. Although this sounds good it might tend to overfitting.

Since your number of inputs and your dataset size is small its ok to start with a hidden layer size of the double (18) and start lowering down. When the training error and test error stabilize in a difference lower than a threshold then you could have found a better generalizing model.

Neural networks are very good at finding local optima by exploring deeply a solution from a starting point. However, the starting point it is also very important. If you are not getting a good generalization you might try to find good initial starting points with methods of Hybrid Neural Networks. A common one, for example, is using genetic algorithms to find an initial combination of weights and then start the neural from that point. Given that your search space would be better covered (in case your problem actually needs that).

As for every problem in machine learning is very important to clean your data before introducing it to the NN. Try to be very detailed in order to avoid the NN to learn things you already know. For example if you know how two features are correlated improve the input data by making this correlation explicit so less workload is given to the NN (that might actually get you in trouble).

• thank you @Javierfdr , I would have a question: Do you suggest the use of genetic algorithm after selecting the number of hidden neurons (and so for start the training with the selected architecture) or for each hidden neuron number trial, in order to have a better starting point before select the hidden neuron number? Dec 25 '14 at 20:37
• You can do both @Daniel, it depends on what can you actually compute. The input optimization with GA will tend to make a specific architecture of NN to work better. So if you can do it for every architecture you are trying it you will have more specific results to analyze. So if your final methodology will be: Optimize initial weights with a GA and use them to a given NN architecture, then yes you will need to do it for each NN to test which works better with this hybrid approach. Please notice that the objective function of the GA will be the calculation of the whole NN for a given input. Dec 26 '14 at 15:42
• Also notice @Daniel that if you are using binary input data you must look for Genetic Algorithms, and if using not binary input data then Evolutionary Algorithms will work better (based in the same principle but better architected for non input data) Dec 26 '14 at 15:45
• Thank you @Javierfdr ! I have performed my test, in the question now there is the result. Can you confirm/comment my deduction? Thanks Dec 26 '14 at 20:43
• @Daniel in general it looks like a good solution. Both 12 neurons and 18 neurons give you a low difference error between the sets. I would chose 12 neurons because it will tend to overfit less than a higher number of neurons. Also, are you training the network long enough? You might use a sort of early stopping to stop learning when a certain threshold of the difference of learnings in two consecutive steps is not surpassed. Look at these nice rule of thumbs for picking number of hidden neurons faqs.org/faqs/ai-faq/neural-nets/part3/section-10.html Dec 26 '14 at 23:30

## Top level:

The rule is to chose the most simple network that can perform satisfactorily. See this publication and it's pdf.

## The Methodology:

So do your proposed test (training many networks at each number of hidden nodes) and plot the results. At the minimum number of nodes, you'll see the worst performance. As you increase the number of nodes, you'll see an increase in performance (reduction of error). At some point N, you'll see the performance seems to hit an upper limit and increasing nodes beyond this will stop giving significant performance gains. Further increases may start to hurt performance a little as training gets more difficult). That point N is the number of nodes you want.

# How it worked for me:

The first time I used this methodology, it created a beautiful almost-sigmoid-looking function with a very clear number of nodes that were needed to achieve good results. I hope this works for you as well as it worked for me.