I'm trying to tackle a multivariate nonlinear regression problem that takes around 20 inputs and outputs around 200.
I have a set of known points and need to come up with a performant neural network able to generate any interpolation points between these known points.
It's not a classification problem, more like a nonlinear multidimensional curve fitting (well, as the first sentence states, it is a multivariate nonlinear regression).
And I was wondering if instead of having a fully connected hidden layer, I modeled a number of parallel dense layers (one for each output, and connected to a single output neuron) and then concatenating the results of those parallel layers to assemble the final output.
Does anyone know the difference (in practice) between the approaches? When should I use each one? Is there an even better approach for such problem?
One thing I realized is that the parallel one reduces the inter-dependency between outputs...