Increasing the learning Capacity of BCI Systems via CNN-HMM models.

Despite all the work in the Brain Computer Interface (BCI) community, one of the main issues that prevents it from becoming pervasive is the limitation on the number of commands with a satisfactory accuracy of detection. In this paper, we propose a solution to increase the number of commands while maintaining a satisfactory accuracy performance via a hybrid Convolutional Neural Network (CNN)- Hidden Markov Model (HMM). The setup makes use of a classifier (a CNN) that works over a small alphabet of established mental tasks like the motor imagery task and detects sequences comprised of these tasks using HMMs. To optimize the learning capacity, we select a subset of sequences by measuring the distance between HMM models. This system, based on the experiments we have conducted, shows a 14% gain in accuracy over the non-sequenced classifier. Alternatively, it can be used to increase the command set size by 4 times when using all the channels or by 1.5 times when using only 1/3 of the EEG channels and have the same performance as a non-sequenced classifier that uses all available channels. This shows that the CNN-HMM hybrid model is a viable approach to increase the capacity of learning in BCI.