Abstract:MOTIVATION:While generative models have shown great success in sampling high-dimensional samples conditional on low-dimensional descriptors (stroke thickness in MNIST, hair color in CelebA, speaker identity in WaveNet), their generation out-of-distribution poses fundamental problems due to the difficulty of learning compact joint distribution across conditions. The canonical example of the conditional variational autoencoder (CVAE), for instance, does not explicitly relate conditions during training and, hence, has no explicit incentive of learning such a compact representation.RESULTS:We overcome the limitation of the CVAE by matching distributions across conditions using maximum mean discrepancy in the decoder layer that follows the bottleneck. This introduces a strong regularization both for reconstructing samples within the same condition and for transforming samples across conditions, resulting in much improved generalization. As this amount to solving a style-transfer problem, we refer to the model as transfer VAE (trVAE). Benchmarking trVAE on high-dimensional image and single-cell RNA-seq, we demonstrate higher robustness and higher accuracy than existing approaches. We also show qualitatively improved predictions by tackling previously problematic minority classes and multiple conditions in the context of cellular perturbation response to treatment and disease based on high-dimensional single-cell gene expression data. For generic tasks, we improve Pearson correlations of high-dimensional estimated means and variances with their ground truths from 0.89 to 0.97 and 0.75 to 0.87, respectively. We further demonstrate that trVAE learns cell-type-specific responses after perturbation and improves the prediction of most cell-type-specific genes by 65%.AVAILABILITY AND IMPLEMENTATION:The trVAE implementation is available via github.com/theislab/trvae. The results of this article can be reproduced via github.com/theislab/trvae_reproducibility.

Conditional Flow Variational Autoencoders for Structured Sequence Prediction

Flow-Based Variational Sequence Autoencoder

PFVAE: A Planar Flow-Based Variational Auto-Encoder Prediction Model for Time Series Data

Flow-based Spatio-Temporal Structured Prediction of Motion Dynamics

FlowSeq: Non-Autoregressive Conditional Sequence Generation with Generative Flow

A Conditional Flow Variational Autoencoder for Controllable Synthesis of Virtual Populations of Anatomy

Causal Flow-based Variational Auto-Encoder for Disentangled Causal Representation Learning

Automatic variational inference with cascading flows

Conditional Variable Flow Matching: Transforming Conditional Densities with Amortized Conditional Optimal Transport

Conditional Out-of-distribution Generation for Unpaired Data Using Transfer VAE.

CauF-VAE: Causal Disentangled Representation Learning with VAE and Causal Flows

Conditional Temporal Variational AutoEncoder for Action Video Prediction

Variational autoencoders with latent high-dimensional steady geometric flows for dynamics

Improving Multimodal Joint Variational Autoencoders through Normalizing Flows and Correlation Analysis

Normalizing Flow with Variational Latent Representation

Recommendation via Collaborative Autoregressive Flows

VAE^2: Preventing Posterior Collapse of Variational Video Predictions in the Wild

NVAE: A Deep Hierarchical Variational Autoencoder

Data-Dependent Conditional Priors for Unsupervised Learning of Multimodal Data

Tions via invertible generative flows

TimeVAE: A Variational Auto-Encoder for Multivariate Time Series Generation