After doing and EDA … ¿it’s a good practice to get rid of the correlated features before feeding in a fully connected NN? … or ¿should I let the NN try to figure out the relationships between the features, being present those correlated? …
Input preprocessing can be useful, but in vision they usually do different things than decorrelation.
In general, I’d say that transformations can help with some hard to train models (you’d basically scale first layer’s gradients to help training), otherwise NN will “figure it out”, especially with adaptive learning rate optimizers.