Deep Learning and Unsupervised Feature Learning

Machine learning algorithms are very sensitive to the representations chosen for the data so it is desirable to improve learning algorithms that can discover good representations, good features, or good explanatory latent variables. Both supervised and unsupervised learning algorithms have been proposed for this purpose, and they can be combined in semi-supervised setups in order to take advantage of vast quantities of unlabeled data. Deep learning algorithms have multiple levels of representation and the number of levels can be selected based on the available data. Great progress has been made in recent years in algorithms, their analysis, and their application both in academic benchmarks and large-scale industrial settings (such as machine vision/object recognition and NLP, including speech recognition). Many interesting open problems also remain, which should stimulate lively discussions among the participants.