Semi-Supervised Learning with Sparse Distributed Representations

For many machine learning applications, labeled data may be very difficult or costly to obtain. For instance in the case of speech analysis, the average annotation time for a one hour telephone conversation transcript is 400 hours.[7] To circumvent this problem, one can use semi-supervised learning algorithms which utilize unlabeled data to improve performance on a supervised learning task. Since unlabeled data is typically much easier to obtain, this can be an attractive approach.