Shift-invariant sparse coding for audio classification Roger Grosse MIT SAIL To apply supervised machine learning to signal classification tasks, one first identifies a set of features which likely summarize the relevant properties of a signal. There are highly effective, general purpose, off-the-shelf algorithms which learn a decision rule for a given set of features, but choosing the features themselves still requires much human effort and ingenuity. The self-taught learning framework aims to automate this part of the process by automatically learning good features from large amounts of unlabeled data. Previous approaches include sparse coding, an algorithm borrowed from computational neuroscience. Features automatically learned with sparse coding have achieved good performance in domains as diverse as visual object recognition and web page classification. Traditional sparse coding learns a representation where a small excerpt from a visual or auditory signal can be modeled as a linear combination of a small number of basis functions drawn from a larger dictionary. Unfortunately, the learned representation is highly redundant, and it is impractical to use sparse coding to represent full-length signals. Shift-invariant sparse coding (SISC) is an extension of sparse coding which solves both problems by allowing the basis functions to be placed in all possible shifts. Our main contribution is an algorithm which performs SISC orders of magnitude faster than previous approaches, making it practical to apply SISC to audio classification tasks. SISC's learned high-level representations of speech and music provide useful features for classification tasks within those domains, suggesting its usefulness as a general-purpose algorithm for learning features for classification tasks.