VAE-SNE: a deep generative model for simultaneous dimensionality reduction and clustering

Abstract

Scientific datasets are growing rapidly in scale and complexity. Consequently, the task of understanding these data to answer scientific questions increasingly requires the use of compression algorithms that reduce dimensionality by combining correlated features and cluster similar observations to summarize large datasets. Here we introduce a method for both dimension reduction and clustering called VAE-SNE (variational autoencoder stochastic neighbor embedding). Our model combines elements from deep learning, probabilistic inference, and manifold learning to produce interpretable compressed representations while also readily scaling to tens-of-millions of observations. Unlike existing methods, VAE-SNE simultaneously compresses high-dimensional data and automatically learns a distribution of clusters within the data — without the need to manually select the number of clusters. This naturally creates a multi-scale representation, which makes it straightforward to generate coarse-grained descriptions for large subsets of related observations and select specific regions of interest for further analysis. VAE-SNE can also quickly and easily embed new samples, detect outliers, and can be optimized with small batches of data, which makes it possible to compress datasets that are otherwise too large to fit into memory. We evaluate VAE-SNE as a general purpose method for dimensionality reduction by applying it to multiple real-world datasets and by comparing its performance with existing methods for dimensionality reduction. We find that VAE-SNE produces high-quality compressed representations with results that are on par with existing nonlinear dimensionality reduction algorithms. As a practical example, we demonstrate how the cluster distribution learned by VAE-SNE can be used for unsupervised action recognition to detect and classify repeated motifs of stereotyped behavior in high-dimensional timeseries data. Finally, we also introduce variants of VAE-SNE for embedding data in polar (spherical) coordinates and for embedding image data from raw pixels. VAE-SNE is a robust, feature-rich, and scalable method with broad applicability to a range of datasets in the life sciences and beyond.