kavuri@lips.ecn.purdue.edu (Surya N Kavuri ) (02/06/91)
I posted some questions on classification. As it is quite shabby, I am writing it again. (Q1) If I have a set of separable classes, am I always guaranteed proper classification using a two layered backprop ? I am not interested in linearly separable cases which are trivial. In particular, what if the classes share a common nonlinear boundary ? Example: I have a circle inside a square. I have two classes: Class I: INSIDE CIRCLE Class II: OUTSIDE CIRCLE The domain of inputs is the square (let it be a unit square). Now, using two inputs (the two coordinates x1 and x2), and one output to identify the class (1 for inside circle 0 else). Certainly I cannot solve this. If I use two output nodes, one for each class, even then I am not sure I can classify. We need at least three lines(hyperpl.) to separate the circle and the two output nodes give only two hyperplanes. Should I expect the sigmoid to some how "bend" the hyperplane ? (Q2) If I use a hidden layer, can I expect the problem to be solvable ? If so, is it because of the increased dimensionality (may be more than 2 hidden nodes) at the hidden layer ? Simply, WHAT ?! SURYA KAVURI (FIAT LUX)