Studying the Decoding CNN Training on Unbalanced Data

The main aim of this study is to prove that network tunning in case of unbalanced data is an actual issue and is the room for the network performance improvement. As a result of conducted research and experiments in that study, hypothesis was actually manifested and statistically proven on simplified simulated case. Moreover, the same results were found for balanced data also, if both the sources are equally prominent, network will still sharpen to only one in some cases. After this, investigation was carried out on parameters influencing the network tuning to uncover how to prevent it. It appeared that hints for spatial filter to highlight more task-related sources from the beginning, have certain effect. Specifically, that tested hint was very simple assumption on initial filter weights, that should a bit separate sources to different network branches, given that we knew topographies of sources as we modeled training data. In a real-world case, such as the 4th BCI Competition IV dataset, information about the electrode positions may be unknown and cues for spatial filter become more problematic to design. Therefore, the main idea for highlighting more sources is to increase amount of output unmixed signals from the spatial filter, assuming that they will require more sources information, and also to divide input sensor data into disjoint groups and project them to different spatial filter branches, so the different branches will more likely catch the different sources. The main outcome is an increased averaged correlation coefficient from 0.45 to 0.48. Further study of the results is necessary to confirm or refute the idea that splitting the input data led to improved correlation, since in reality, the architecture change is simply a modification of the same operation, and the real reason for the performance improvement may be completely different.