Recovering data from high amounts of loss and corruption would be useful for a wide variety of civilian and military applications. Highly corrupted data (e.g., speech and images) has been less studied relative to the problem of light corruption, but would be advantageous for applications such as low-light imagery and weak signal reception in acoustic sensing and radio communication. Unlike milder signal corruptions, resolving strong noise interference may require a more robust approach than simply removing predictable noise, namely actively looking for the expected signal, a type of problem well suited for machine learning. In this work, we evaluate a variant of the U-net autoencoder neural network topology for accomplishing the difficult task of denoising highly corrupted images and English speech when noise floors are 2-10x stronger than the clean signal. We test our methods on corruptions including additive white Gaussian noise and channel dropout.
Gaining insight from unlabeled data is a widespread, challenging problem with many immediate applications. Clustering, dimensionality reduction for visualization, and anomaly detection are unsupervised learning solutions to this problem. Typically, the efficacy of these methods relies on obtaining sufficient amounts of data, presenting many challenges to cases where only limited data exists. In this work, we demonstrate that deep convolutional autoencoders can comfortably perform these tasks either directly or through manipulations of the latent space in a limited data setting. Clustering on common benchmark datasets produces comparable results to the current state-of-the-art in unsupervised classification. Clustering the activation maps of the encoding layers results in a form of unsupervised image segmentation with limited two-dimensional data. Visualizing the activation maps through dimension reduction demonstrates the possibilities of anomaly detection and semi-supervised learning. We are currently utilizing our versatile autoencoder to explore the ambitious task of finding anomalous and/or inconspicuous objects from single images.