Using Machine Learning to Recreate Signals from the Primary Visual Cortex of Mice

Abstract

Generative adversarial networks (GAN) have received much attention lately for its use with images and has been shown to be able to create extremely realistic images of different kinds of objects. Even though its use in images is popular, there has not been much study into using this type of generative method for other types of data. GANs replicate the distribution of a data set to produce realistic samples that are not in the data set. Being an adversarial method based on game theory, training a GAN can be difficult. If not tweaked correctly, the GAN can collapse and produce unrealistic results. A different kind of machine learning method called an autoencoder is more stable as it is based on replicating data instead of replicating the distribution of the data. Experimental biology needs large amounts of data, which is sometimes difficult to procure in sufficient amounts. Using generative methods in these instances have much potential. Proposed here is using an autoencoder to stabilize training of a GAN for use with biological signals. Autoencoders were shown to be stable, and replicated input data almost exactly. The GAN model was shown to perform better when pre-trained with an autoencoder. The samples produced by the GAN were not realistic as the model requires more training.