The method of “Deep Learning” has gained a lot of interest due to recent impressive successes in areas such as image processing, speech recognition and artificial intelligence/ human-machine interaction (e.g. Go playing), where deep learners have consistently outperformed previous state-of-the-art approaches. Deep learners use multiple layers of non- linear functions to learn mappings from input to output data, e.g. from music audio signals to semantic descriptions thereof like genre, mood or other semantic tags. Deep learners are probably the most prospering artificial intelligence method at the moment.
Given the tremendous success of deep learning, there is, however, surprisingly little systematic knowledge and formal understanding of its principles of operation. Formulation of mathematical theory explaining the achievements of deep learning has just now begun. In particular, recently developed methods from applied harmonic analysis are able to address and explain aspects of the structure of deep learning algorithms.
This workshop will assemble scientists from both applied mathematics and computer science, who are working on the aforementioned problems and are, in particular, interested in gaining a systematic formal understanding of the underlying structures and principles of deep learning. Since the organizers’ project “SALSA” is focused on music and audio processing, but also, because this application area has so far obtained less attention than, e.g., image processing, we are planning to emphasize the application of deep learning methods to audio.