I want to use a neural network to predict the refractive index of a solution. My thinking is, instead of immediately training on many samples, I will first find the 'ultimate resolution' of the network given the experimental apparatus I am using. What I mean is I will make two different solutions which have refractive indices near the middle of the range of which I am interested. Then I will train the network to classify these two solutions based on reflectance measured from the solution. If it works, say with at least 95% accuracy, then I will make two different solutions in which the difference in refractive index is smaller than before. I will repeat this until the ANN classifies, say below 95%.
Will this method of finding the 'resolution' by classification extrapolate well to regression with many more training examples?