A convolutional neural network for generalized and efficient spike classification

In neuronal recordings, analysis can classify electrode waveforms into spikes and noise. However, many automated sorting algorithms are highly variable in classification across different recordings and different implant areas. Here we trained a Convolutional Neural Network (CNN) on prelabeled waveforms collected from in vivo cortical recordings. The network, once trained, outputs a likelihood value that an input waveform should be a spike. To compare our network, we used a previous design from our lab using only fully connected layers, making a case for the benefit of convolutional layers for spike classification. We also compared classification across multiple cortical areas, showing improvement in classification accuracy and sensitivity to threshold parameters. Our classifier serves as a robust preprocessing step that can be applied to a diverse array of waveforms with predictions similar to that of a human sorter.