On Disentangled Representations Learned From Correlated Data

Abstract

The focus of disentanglement approaches has been on identifying independent factors of variation in data. However, the causal variables underlying real-world observations are often not statistically independent. In this work, we bridge the gap to real-world scenarios by analyzing the behavior of the most prominent disentanglement approaches on correlated data in a large-scale empirical study (including 4260 models). We show and quantify that systematically induced correlations in the dataset are being learned and reflected in the latent representations, which has implications for downstream applications of disentanglement such as fairness. We also demonstrate how to resolve these latent correlations, either using weak supervision during training or by post-hoc correcting a pre-trained model with a small number of labels.

arXiv commentPublished at the 38th International Conference on Machine Learning (ICML 2021)

References in corpus (19)

• Auto-Encoding Variational Bayes
• Stochastic Backpropagation and Approximate Inference in Deep Generative Models
• Shortcut Learning in Deep Neural Networks
• Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations
• Disentangling by Factorising
• Unsupervised Feature Extraction by Time-Contrastive Learning and Nonlinear ICA
• Multi-Level Variational Autoencoder: Learning Disentangled Representations from Grouped Observations
• Flexibly Fair Representation Learning by Disentanglement
• Are Disentangled Representations Helpful for Abstract Visual Reasoning?
• Weakly-Supervised Disentanglement Without Compromises
• SCAN: Learning Hierarchical Compositional Visual Concepts
• Weakly Supervised Disentanglement with Guarantees
• Bias and Generalization in Deep Generative Models: An Empirical Study
• Learning Neural Causal Models from Unknown Interventions
• CausalVAE: Structured Causal Disentanglement in Variational Autoencoder
• Disentangled State Space Representations
• Exact Rate-Distortion in Autoencoders via Echo Noise
• Evaluating the Disentanglement of Deep Generative Models through Manifold Topology
• Spatial Dependency Networks: Neural Layers for Improved Generative Image Modeling

Cited by in corpus (11)

• Learning Disentangled Representations in the Imaging Domain
• Learning Robust Representation for Joint Grading of Ophthalmic Diseases via Adaptive Curriculum and Feature Disentanglement
• Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding
• Desiderata for Representation Learning: A Causal Perspective
• On the Transfer of Disentangled Representations in Realistic Settings
• Benchmarks, Algorithms, and Metrics for Hierarchical Disentanglement
• Visual Representation Learning Does Not Generalize Strongly Within the Same Domain
• Generalization and Robustness Implications in Object-Centric Learning
• Disentangling representations of retinal images with generative models
• Factorized Gaussian Process Variational Autoencoders
• DEFT: Distilling Entangled Factors by Preventing Information Diffusion