[Paper Review] Domain Agnostic Learning with Disentangled Representations

Xingchao Peng, Zijun Huang+2 more|arXiv (Cornell University)|Apr 28, 2019

Domain Adaptation and Few-Shot LearningComputer Science137 citations

TL;DR

The paper introduces Domain-Agnostic Learning (DAL) and a Deep Adversarial Disentangled Autoencoder (DADA) to disentangle domain-invariant, domain-specific, and class-irrelevant features, enabling knowledge transfer from a labeled source to unlabeled, mixed-target domains and achieving state-of-the-art results on several datasets.

ABSTRACT

Unsupervised model transfer has the potential to greatly improve the generalizability of deep models to novel domains. Yet the current literature assumes that the separation of target data into distinct domains is known as a priori. In this paper, we propose the task of Domain-Agnostic Learning (DAL): How to transfer knowledge from a labeled source domain to unlabeled data from arbitrary target domains? To tackle this problem, we devise a novel Deep Adversarial Disentangled Autoencoder (DADA) capable of disentangling domain-specific features from class identity. We demonstrate experimentally that when the target domain labels are unknown, DADA leads to state-of-the-art performance on several image classification datasets.

Motivation & Objective

• Address domain shift where target data come from multiple unlabeled domains without domain labels.
• Learn domain-invariant features while removing domain-specific and class-irrelevant information.
• Enhance disentanglement with mutual information minimization and adversarial training.
• Demonstrate that class disentanglement boosts domain adaptation performance.
• Show state-of-the-art results on standard image classification benchmarks under the DAL protocol.

Proposed method

• Introduce the Deep Adversarial Disentangled Autoencoder (DADA) that splits features into domain-invariant, domain-specific, and class-irrelevant components.
• Use a Variational Autoencoder (VAE)-based encoder/decoder framework for disentanglement and reconstruction.
• Apply class disentanglement via an adversarial setup with a class identifier trained on the source to predict labels, while the disentangler fools it (L_ent).
• Apply domain disentanglement with a domain identifier to distinguish source vs. target in the latent space, training the disentangler to produce domain-invariant features (L_DI).
• Minimize mutual information between (domain-invariant, domain-specific) and (domain-invariant, class-irrelevant) features using Mutual Information Neural Estimator (MINE).
• Incorporate ring-style normalization to stabilize embeddings across heterogeneous domains (ring loss and GM variant).
• Train end-to-end with alternating objectives including reconstruction (L_vae) and adversarial/disentanglement losses; use a reconstruction mechanism to recover G from (f_di, f_ds) or (f_di, f_ci).

Experimental results

Research questions

• RQ1 Can knowledge be transferred from a single labeled source domain to unlabeled, heterogeneous target domains without domain labels?
• RQ2 Does disentangling domain-invariant features from domain-specific and class-irrelevant features improve transfer performance?
• RQ3 Does adversarial class disentanglement combined with mutual information minimization yield superior domain-agnostic representations?
• RQ4 How do ablated components (class disentanglement, domain disentanglement, ring loss, reconstruction) affect performance?
• RQ5 What is the impact of DADA on standard DAL benchmarks like Digit-Five, Office-Caltech10, and DomainNet?

Key findings

• DADA achieves state-of-the-art performance on several domain-agnostic learning tasks across Digit-Five, Office-Caltech10, and DomainNet.
• Ablation studies show that both class disentanglement and domain disentanglement contribute to performance gains, with the full model (including ring loss and reconstruction) providing the best results.
• Incorporating ring-style normalization (ring loss) significantly improves performance, demonstrating the importance of stable feature normalization in DAL.
• Mutual information minimization via MINE strengthens disentanglement and improves alignment between source and heterogeneous targets, reflected in empirical results.
• t-SNE visualizations indicate that DADA features yield clearer class separation than baselines like UFDN and MCD, supporting better discriminability in the target domains.
• Quantitative analysis shows that DADA reduces domain discrepancy (A-distance) in the learned features compared to baselines, aligning source and target distributions more effectively.