Applicability of echo state networks to classify EEG data from a movement task

Abstract

Prosthetic devices have come a far way from being just mechanical devices. In recent years, neuroprosthetic devices have been developed, that directly infer movements commands from neuronal activities. Amongst these, hand prostheses require a more precise detection of different hand motions than other body parts. However, detection of such precise movements in EEG data is a non-trivial task due to the noisiness. To challenge this problem, the WAY consortium created a classification challenge on the Kaggle platform in the summer 2015. The winners used a recurrent convolutional neural network that scored 0.98 AUC. Since training such a network is computationally demanding, we applied an echo state network to the same dataset, to see whether this faster approach can compete with the RCNN. The dataset originates from grasp and lift trials recorded by the WAY consortium [1]. They labeled the data with six different events, occurring in the same order for each trial. Further, each of the events are labeled ±75 ms around the onset of the event. Lastly, the dataset is imbalanced, as most of the time no event occurs. To challenge this imbalanced, we used a weighted ridge regression to learn the weights of the output layer. We further tried subsampling the frames where no event occurred. Preliminary results suggest no significant difference between these two methods. Additionally, we used different activation functions including hyperbolic tangent and rectified hyperbolic tangent. Lastly, we set up the reservoir in three “bubbles” that were highly connected, whilst between bubbles only few connections were active. Other than that, the original approach of Jaeger was used [2]. To preprocess the data, we used common spatial patterns. We applied these to each event separately against the rest time, leading to six different preprocessed datasets. For each of these datasets, we classified the corresponding event using one ESN. Afterwards, we concatenated the predictions to form the original six events. We have not yet fully evaluated this approach. However, preliminary results (ca. 0.76 AUC) are promising, although they do not compete with the results of the competition winners. © Springer International Publishing Switzerland 2016.