Removing noise and overlapping spikes from extracellular recordings using a regularized denoising autoencoder

Abstract

Abstract One of the difficulties in sorting extracellularly recorded neuronal action potentials, known as spikes, is noise from several sources superimposed on the target spike. For example, the activity of other neurons is often overlaid on the recorded spikes, denoted as neuronal noise interference (NNI), and interference from neighbouring neurons firing at the same time is known as overlapping spikes. In addition, the noise induced by the recording device is known as Gaussian white noise (GWN). Thus, the appropriate signal denoising process of extracellular recordings is an extremely challenging task, where the classical filtering approach may fail specially in removing the overlapping spikes due to the similarity between the spike waveforms of the neighbouring neurons and the wanted single-unit waveform. This issue is tackled in this work by training regularized denoising autoencoder (RDAE) model with the average spike-waveforms of previously identified single-units, denoted as ground truth. To capture the distinct features of the wanted single-unit waveforms, and weight regularization were employed at the encoder/decoder of the RDAE, respectively. With a suitable regularization parameter, the -RDAE model outperforms the -RDAE model. In general, RDAE shows extraordinary performance not only in removing the NNI and GWN, but also the overlapping spikes.