Learning to Detect Faces with Snow
A novel learning approach for face detection in still images using a network of linear units is presented. The SNoW learning architecture is a sparse network of linear functions over a pre-defined or incrementally learned feature space and is specifically tailored for learning in the presence of a very large number of features. A wide range of face images in different poses, with different expressions and under different lighting conditions are used as a training set to capture the variations of human faces. Experimental results on commonly used CMU data sets show that the SNoW-based approach perform well against methods that use neural networksneural network, Bayesian classifiers, Support Vector Machines and others. To quantify and explain the experimental results, we present a theoretical analysis that shows the advantage of this architecture and traces it to the nature of the update rule used in SNoW, a multiplicative update rule based on the Winnow learning algorithm. In particular, in sparse domains (in which the number of irrelevant features is large) this update rule is shown to be advantageous relative to algorithms that are derived from additive update rules such as Perceptron and Support Vector Machines. We show that learning problems in the visual domain have these sparseness characteristics and exhibit it by analyzing data taken from face detection experiments. Our experiments exhibit good generalization and robustness properties of the SNoW-based method, and conform to the theoretical analysis.