Many subfields of political science estimate ideal points of politicians, parties, justices, and voters based on these actors' observed decisions. We develop a novel, flexible approach to estimating ideal point models via neural networks. This approach enables the derivation of ideal points from new types of unstructured data (e.g., text, images, videos). Specifically, we use a regularized autoencoder architecture to approximate the distribution of ideal points conditional on a researcher prior. The model displays good finite sample performance in Monte Carlo simulations and produces near-identical estimates compared to well-established models in the literature on common datasets (e.g. W-Nominates, IRT). It accommodates all data types, all forms of covariates, any mapping between ideal points and observed decisions, and any Bayesian prior on the distribution of ideal points. It shows its true potential when combined with state-of-the-art embeddings to estimate ideal points on massive multilingual and multimodal datasets. A Python package, IdealPointNN, is provided to support future applications.