An Anticipatory Self-Organized Map for Robust Recognition

When performing any real-time detection task, such as face detection, speech recognition, etc., we can take advantage of the temporal correlations within the data stream. This can help us make detection more robust by using anticipations about the target to overcome the variance due to noise. We present an extended self-organized map that uses lateral weights between the nodes to learn temporal relations between clusters. These weights are then used during recognition to bias certain nodes to win the competition. This converts the self-organized map from a maximum likelihood to a maximum a posteriori estimator. We present an experiment using artificial data to demonstrate the benefit of the anticipatory self-organized map.