Unsupervised Feature Learning Via Sparse Hierarchical Representations

Book excerpt: Machine learning has proved a powerful tool for artificial intelligence and data mining problems. However, its success has usually relied on having a good feature representation of the data, and having a poor representation can severely limit the performance of learning algorithms. These feature representations are often hand-designed, require significant amounts of domain knowledge and human labor, and do not generalize well to new domains. To address these issues, I will present machine learning algorithms that can automatically learn good feature representations from unlabeled data in various domains, such as images, audio, text, and robotic sensors. Specifically, I will first describe how efficient sparse coding algorithms --- which represent each input example using a small number of basis vectors --- can be used to learn good low-level representations from unlabeled data. I also show that this gives feature representations that yield improved performance in many machine learning tasks. In addition, building on the deep learning framework, I will present two new algorithms, sparse deep belief networks and convolutional deep belief networks, for building more complex, hierarchical representations, in which more complex features are automatically learned as a composition of simpler ones. When applied to images, this method automatically learns features that correspond to objects and decompositions of objects into object-parts. These features often lead to performance competitive with or better than highly hand-engineered computer vision algorithms in object recognition and segmentation tasks. Further, the same algorithm can be used to learn feature representations from audio data. In particular, the learned features yield improved performance over state-of-the-art methods in several speech recognition tasks.