Learning multiple categories from sequences of examples

We propose a neural network architecture together with a new learning algorithm to learn representations of multiple categories. More specifically, the algorithm learns in a supervised manner from sequences of examples of each category. We will show that our algorithm approximates the minimum of a quadratic homogeneous program. This minimum has a natural interpretation, it separates each category maximally from the mean of all the other categories. Finally, we show some examples of how our algorithm works.