I have heard a colleague of mine giving the following statements to a student, but I am not quite sure if he is right. The statements were about Multi Layer Perceptron and the Backpropagation algorithm, he said that:
It would be reasonable to check it up if the data to be input to a MLP should be checked first its degree of data normality. It is the first time I heard that and I am not sure about it. I have seen this article: https://www.researchgate.net/publication/256980936_Assumptions_of_Multiple_Regression_Correcting_Two_Misconceptions
that talks about multiple regression (and I know that a MLP NN can be used for this case) and it says that:
In this article, we clarify that multiple regression models estimated using ordinary least squares require the assumption of normally distributed errors in order for trustworthy inferences, at least in small samples, but not the assumption of normally distributed response or predictor variables
So a MLP with BP algorithm does it only work when the input data if it has a normal distribution?
Another point that he mentions is that a MLP NN could not work in n-dimensions, but I am pretty sure that is not a limitation, because I have seen in the Andrew Ng course that he talks about data that is in Rn dimensions, and uses MLP and backpropagation to classify it. So both limitations could exist? in which cases a MLP could fail, regarding the data input.