Representation Learning for Feature Analysis in Synthetic Neuronal Signals

The current standard for neural signal spike detection uses an amplitude-thresholding approach, which is biased towards high-amplitude signals and evoked signals with artifacts. Because detection is an initial step in signal analysis, it is critical to have an accurate representation of spike occurrence and shift the detection paradigm to focus on the signal shape. One limitation of the current approach is that it tends to disregard low-amplitude signals, which carry important information such as sensory and pain response. Thus, improvements in understanding the shapes and characteristics of neural signal features is vital to enhance signal analysis quality and neuromodulation techniques. We propose the use of a generative autoencoder model to learn a representation of neuronal signal characteristics using an unsupervised training approach. In this work, a synthetic neural signal dataset is constructed using Multi-Electrode-Array recordings (MEArec) software. The synthetic data is used to train an autoencoder, a model which learns a latent representation of the input data, and outputs a reconstructed version of it. The accuracy of the autoencoder in recreating the input data reflects the quality of its latent representation. The use of synthetic data allows for the direct correlation of latent representation learning and validation loss with ground truth neuronal features. We hypothesize that by learning a robust representation of data with known ground truths, the latent space of the autoencoder will contain salient information on neuronal feature characteristics. In future work, this learned representation will be leveraged to detect and identify salient features within experimental data.