This project on a new perspective on selective attention, aiming at suggesting a technique
grounded on the model-based learning (model fitting) to select the attended voice (AV) from multiple sound streams, started last year and we aim at running it for one year more. This project is a collaboration between Linköping University and Eriksholm (Independent Research center of the hearing aid manufacturer Oticon). The long-term goal of this project would be to evaluate this technique on hearing impaired persons, and to develop a prosthesis that amplifies only AV. Here we focus on deep learning and deep neural networks (DNNs) applied on EEG and audio data, and in the later stage of the project, on unobtrusive wearable EEG (EarEEG) data, that is EEG collected in the ear canal. We aim at designing design a model with high performance using advanced machine learning algorithms, what requires an extensive experimentation with data utilities ranging from data collection and storage (in huge volumes) to data analysis.
Our preliminary assumptions on DNN network design and requirements are as follows. We
aim at using RNN with GRU/LSTM cells. Our planned is architecture (some assumptions):
(1) model composed of attentional encoder-decoder model, (2) 4 hidden layers, and (3) 1000
hidden nodes per layer. To speed up the training, we would like to have each layer on different GPUs. Therefore we are seeking Kebnekaise system at HPC2N in Umea.