The brain-computer interface systems utilize modern scientific results to unravel the yet unknown secrets of the brain and also to ensure medical and other lucrative applications. The basis of these extensively used systems is the recognition of single neuron activities.
Reliable spike detection and sorting methods are already available in practice, but for further improvements, it is advisable to utilize not just the temporal but also the spatial information of the multi-channel neural probe.
This paper introduces an artificial neural network based system, which was designed to be able to recognize action potential waveforms by putting a great emphasis on the spatial connectivities. I also present a graphical user interface created to provide a convenient visualization form for the multidimensional time series. The graphical representation supports the network evaluation process too. The convolutional neural network was trained on a 128-channel measurement recording and it gained the ability to extract spatial features from its input. With an additional LSTM layer, the system became able to precisely detect the action potential waveforms in the raw data.
The final network will form the basis of future work, which aims to construct a fully automated neural activity detector. Additional optimization techniques can be used to make the multi-channel recordings more informative without supplementary hardware developments.