I am trying to build predictive models for 2 outputs given approximately 10 input variables.

I understand that I would first look at the simplest model, such as a multivariate linear regression, then to build neural networks (e.g. 1 hidden layer) and progressively move on to deeper hierarchies.

I have 2 questions:

1) Consider the 3 neural network architectures below: Each model attempts to predict 2 outputs given a bunch of predictors (input, independent variables). Which would be ideal?

enter image description here

2) How do I decide on which is the best neural network? I have currently performed NetTrain[] on my data and am curious how to evaluate performance metrics within Mathematica.

Appreciate your advice.


  • $\begingroup$ You already mentioned yourself that with only 10 inputs and 2 outputs (that I assume are scalar values) your problem is not a good fit for neural networks. If you really really really want to use a neural network use the simplest archicture possible, use regularization and you better have a large amount of training data as weil $\endgroup$
    – Sascha
    Aug 4, 2017 at 5:53
  • $\begingroup$ Please give a dateset and net code to test. $\endgroup$ Aug 4, 2017 at 6:23

1 Answer 1


Consider the 3 neural network architectures below. Which would be ideal?

That obviously depends on the problem, but the standard approach would be (A), with the reasoning that features learned in lower layers useful for predicting output 1 are probably also useful for predicting output 2. So learning output 2 can improve the quality of output 1 and vice versa.

I don't see the difference between B and C. It seems both have exactly the same connections between input and output.

How do I decide on which is the best neural network?

Normally by using a validation set (NetTrain can do that for you) to see how well it generalizes to unknown data.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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