-1
$\begingroup$

I want to make a simulation based on neural network that will estimate the situation label(not a discrete value) based on state values. Suppose I have data with 40 features/columns and one feature is limited in the range of 25-50, now the goal is the following, when simulating that environment I need to test some states where the value of that limited feature is out of that specified range. I need to know how such a simulation would behave, compared to reality. My initial thought was that the network finds the patterns between features and label, and if I give it an unusual feature value it can still estimate the label pretty accurately.

$\endgroup$

2 Answers 2

0
$\begingroup$

I agree with the conclusion you included within your question. The inherent nature of neural networks - along with the fact that all features interact with each other and the associated weights - will help you in cases where you get "outlying" data. It may even turn out that with 40 features, this one particular feature may not even matter at all regardless of value.

So, I think you are OK to proceed as you stated in your original question.

$\endgroup$
2
  • $\begingroup$ I thinking about taking a data for NN and limit one of it's features in some range, and after training on that data predict on those out of range data points to see if the error is too big. b.t.w. do you, by any chance, have any recommendations of data?? $\endgroup$ Commented Nov 29, 2018 at 7:47
  • $\begingroup$ I think neural networks may help as well but training with a data that you won't have the samples that may occur in the test stage is dangerous. $\endgroup$
    – Ugur MULUK
    Commented Nov 29, 2018 at 11:11
0
$\begingroup$

The values that are far out of the core distribution are called as outliers. In general, we do not want to have outliers in the features since we want generalization. Outliers are the points that are found rarely but increase the error inproportional to the general distribution of your samples. Outliers strongly pull the model through themselves upon training since they create large errors where the optimization task is to minimize it.

If you do not have outliers on your dataset at the training set, do not expect your model to perform well on those points since it has no experience on those. If you have outliers in your training dataset, your model will lose some generalization (it will perform worse in overall), yet your model will perform better at outliers. Your model can only learn what you give at it, neither classic machine learning algorithms nor neural networks are magicians.

Hope I could help. If not, I will be around.

$\endgroup$
3
  • $\begingroup$ even if only one feature is limited in the range?? $\endgroup$ Commented Nov 28, 2018 at 13:38
  • $\begingroup$ I disagree with this answer. Your answer might make sense with something like linear regression, but with neural networks it's a different story. The interactions of each feature, with every other feature, along with the associated weights should help in those cases where you have a feature with an "out of range" value. $\endgroup$ Commented Nov 28, 2018 at 21:56
  • $\begingroup$ The interactions and crosses between features is what neural networks does it best. But think of just one of the features, which has its all distribution in [0,20]. If the value at the real-time comes as 1000+, that could harm your entire net since you have "crosses", you dont use just one feature unrelated to others, like linear regression. Neural networks are robust for those situations in general but testing with a data from a different distribution... $\endgroup$
    – Ugur MULUK
    Commented Nov 29, 2018 at 11:09

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