Hi, Now that I got my ANN training working (thanks to NickGeorgia for help), I started thinking about some theoretical, high-level questions. Regarding neural network classification: Suppose we do some standard ANN classification, e.g try to recognize digits from 0-9 with an ANN. Now, the input of the algorithm is a bitmap, and the output is a classification of the input: the digit whose ANN activation value was the highest. Now, this poses a problem: what if the ANN is fed something like letter "A"? That is, an input which is not designed to be a digit in the first place. How can the ANN tell in this case that "hey dude, this is no digit at all, take it to another ANN"? It seems to me that this is a somewhat hard case. Certainly, we could add another output unit to the ANN and call it "error unit", which all clearly erroneous inputs would activate. But it's really hard to generate training data for this unit! If use totally random bitmaps, we'll just teach it noise samples, and a non-noisy input such as a clear letter "A" will more probably active some of the "legal" 0-9 units that the "noise error unit". Are there any solutions to this problem in the ANN community? Thanks for your time, -- Mikko

You have values that represent how close the input is to each output. Why not consider the input bad if all of those values are too low. If all the outputs are really low you know you have an error.