Comparing Machine Learning and Deep Learning for Octave Illusion Classification Using MEG Data

抄録

The octave illusion is an auditory phenomenon that occurs when two tones with one-octave difference are simultaneously played to both ears repeatedly. This paper aimed to find the most efficient way to classify participants into illusion (ILL) and non-illusion (non-ILL) groups by comparing the amplitude of ASSR at the auditory cortex for the ILL and non-ILL groups using brain data recorded with magnetoencephalography (MEG) among machine learning and deep learning techniques. We used three methods: support vector machine, convolutional neural network, and ensembling neural network for executing data's features. Despite longer training time and less accurate classification results, which could be the result of hyperparameter choice, we believe that ensembling convolutional neural networks is the most efficient way for classification ILL and non-ILL data.