The seminar will introduce and show how machine learning can be applied for time series analysis and prediction. It will start by a general introduction to the topic by Jim T?rresen including recent advances in deep neural networks (DNNs) and other machine learning techniques.
Charles P. Martin will then provide a deep dive on recurrent neural networks (RNNs) and “long short-term memory” units (LSTM). RNNs enable neural networks to be trained with parameters that are shared in between an ordered sequence of tasks. LSTM units introduce a kind of “memory” that can accumulate data over multiple activations, and (most importantly) learn to forget data at the right time. These two technologies have shown much recent promise in sequence learning, particularly for creative tasks such as text and music generation. The technical foundations will be described, as will current ideas for extension of these models to develop future “temporal” AI systems.
Both presentations will also contain examples of our own research, including learning and predicting human behaviour, and “embodied” approaches to musical AI.
The RITMO Seminar Series is a monthly forum for sharing new research.