Making Music Composing Easier for Amateurs: A Hybrid Machine Learning Approach

Learning to store information over extended time intervals by recurrent backpropagation takes a very long time, mostly because of insufficient, decaying error backflow. We briefly review Hochreiter's (1991) analysis of this problem, then address it by introducing a novel, efficient, gradient-based method called long short-term memory (LSTM). Truncating the gradient where this does not do harm, LSTM can learn to bridge minimal time lags in excess of 1000 discrete-time steps by enforcing constant error flow through constant error carousels within special units. Multiplicative gate units learn to open and close access to the constant error flow. LSTM is local in space and time; its computational complexity per time step and weight is O(1). Our experiments with artificial data involve local, distributed, real-valued, and noisy pattern representations. In comparisons with real-time recurrent learning, back propagation through time, recurrent cascade correlation, Elman nets, and neural sequence chunking, LSTM leads to many more successful runs, and learns much faster. LSTM also solves complex, artificial long-time-lag tasks that have never been solved by previous recurrent network algorithms.

Near-infrared spectroscopy (NIRS) has recently been investigated as a non-invasive brain-computer interface (BCI). In particular, previous research has shown that NIRS signals recorded from the motor cortex during left- and right-hand imagery can be distinguished, providing a basis for a two-choice NIRS-BCI. In this study, we investigated the feasibility of an alternative two-choice NIRS-BCI paradigm based on the classification of prefrontal activity due to two cognitive tasks, specifically mental arithmetic and music imagery. Deploying a dual-wavelength frequency domain near-infrared spectrometer, we interrogated nine sites around the frontopolar locations (International 10-20 System) while ten able-bodied adults performed mental arithmetic and music imagery within a synchronous shape-matching paradigm. With the 18 filtered AC signals, we created task- and subject-specific maximum likelihood classifiers using hidden Markov models. Mental arithmetic and music imagery were classified with an average accuracy of 77.2% +/- 7.0 across participants, with all participants significantly exceeding chance accuracies. The results suggest the potential of a two-choice NIRS-BCI based on cognitive rather than motor tasks.