Interactive Visual Exploration of Latent Space (IVELS) for peptide auto-encoder model selection

Abstract

We present a tool for Interactive Visual Exploration of Latent Space (IVELS) for model selection. Evaluating generative models of discrete sequences from a continuous latent space is a challenging problem, since their optimization involves multiple competing objective terms. We introduce a model-selection pipeline to compare and filter models throughout consecutive stages of more complex and expensive metrics. We present the pipeline in an interactive visual tool to enable the exploration of the metrics, analysis of the learned latent space, and selection of the best model for a given task. We focus specifically on the variational auto-encoder family in a case study of modeling peptide sequences, which are short sequences of amino acids. This task is especially interesting due to the presence of multiple attributes we want to model. We demonstrate how an interactive visual comparison can assist in evaluating how well an unsupervised auto-encoder meaningfully captures the attributes of interest in its latent space.