Source-linked AI summary
Domain Agnostic Learning with Disentangled Representations
Xingchao Peng, Zijun Huang, Ximeng Sun, Kate Saenko
TL;DR
The paper addresses transfer from one labeled source to unlabeled target data drawn from arbitrary, mixed domains, challenging methods that assume known target-domain separation. It proposes DADA, an adversarial disentangled autoencoder that separates domain-invariant features from domain-specific and class-irrelevant information while minimizing mutual information between disentangled features. Across DAL image-recognition benchmarks, the authors report that DADA outperforms several state-of-the-art domain-adaptation baselines.
Problem
Domain-Agnostic Learning transfers knowledge from one labeled source domain to multiple unlabeled target domains without target-domain labels, unlike conventional methods requiring separable target domains.
Method
DADA uses adversarial domain and class disentanglement plus mutual-information minimization to learn domain-invariant features separated from domain-specific and class-irrelevant information.
Results
DADA’s Model IV achieves 62.3% average accuracy on Digit-Five and outperforms other baselines on most domain-agnostic tasks.
Takeaways & Limitations
The experiments support disentangled representations, including class disentanglement and mutual-information minimization, for domain-agnostic image recognition.
Abstract
from arXiv · showhide
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.
1. Introduction
The paper introduces Domain-Agnostic Learning for transferring knowledge from one labeled source to mixed, unlabeled target domains. It proposes DADA, which disentangles domain-invariant features from domain-specific and class-irrelevant information to address alignment and negative-transfer challenges.
- Problem: Domain-Agnostic Learning transfers knowledge from one labeled source domain to multiple unlabeled target domains whose samples have mixed domains.This setting removes the assumption that target samples can be isolated into separately labeled homogeneous domains.
- Problem: Mixed target domains hamper mainstream feature-alignment methods designed for one-to-one alignment.The paper identifies unobserved target-domain structure as a central challenge for domain-agnostic transfer.
- Problem: Class-irrelevant information can cause negative transfer, especially when target domains are highly heterogeneous.The paper therefore motivates disentangling class-specific features from remaining image information.
- Approach: DADA disentangles domain-invariant features from both domain-specific and class-irrelevant features simultaneously.Its class disentanglement is trained adversarially using a source-domain class identifier that the disentangler attempts to fool.
- Approach: DADA minimizes mutual information between disentangled features to strengthen their separation.The paper presents this minimizer as a complement to adversarial domain and class disentanglement.
2. Related Work
Prior work addresses domain adaptation, representation disentanglement, and agnostic learning through alignment, generative representations, or normalization. The paper positions DADA as a method for heterogeneous target domains that combines domain and class disentanglement rather than relying on normalization alone.
- Domain Adaptation: Unsupervised domain adaptation methods commonly transfer from labeled source domains to unlabeled target domains using discrepancy, adversarial, or reconstruction-based approaches.These approaches are typically designed for single-source to single-target adaptation.
- Domain Adaptation: Multi-source domain adaptation considers multiple training sources but assumes a single target domain.This differs from the paper’s setting of multiple heterogeneous target domains.
- Representation Disentanglement: Representation-disentanglement research models factors of data variation using approaches including GANs and VAEs.The cited work includes attempts to separate domain-invariant content from domain-specific information.
- Agnostic Learning: Prior agnostic-learning studies include MAML, frameworks for extending knowledge across source domains, and domain-agnostic normalization.The paper distinguishes its focus on transferring knowledge to heterogeneous domains from MAML’s few-shot task adaptation.
- Positioning DADA: DADA’s architecture combines domain disentanglement, class disentanglement, and mutual-information minimization, with adversarial training for the disentanglement components.The class branch removes class-irrelevant features while the domain branch extracts domain-invariant features.
3. DADA: Deep Adversarial Disentangled Autoencoder
DADA addresses domain-agnostic learning by disentangling feature representations into domain-invariant, domain-specific, and class-irrelevant components. Its training combines adversarial class and domain disentanglement, mutual-information minimization, reconstruction, and normalization mechanisms.
- Model architecture: DADA maps each input to domain-invariant, domain-specific, and class-irrelevant features, then reconstructs the original feature representation from disentangled components.The feature generator produces an entangled vector, while the disentangler separates it and the reconstructor recovers the original representation.
- Class disentanglement: Class disentanglement adversarially removes class-irrelevant information by training a class identifier on labeled source data and fooling it with the corresponding features.The paper identifies background as an example of class-irrelevant information.
- Domain disentanglement: Domain disentanglement uses adversarial domain classification so the disentangler extracts domain-invariant features for aligning source and heterogeneous target data.The domain identifier is trained adversarially against the disentangler.
- Mutual information minimization: DADA minimizes mutual information between domain-invariant features and both domain-specific and class-irrelevant features, using MINE to estimate mutual information.MINE uses a neural network estimator and Monte-Carlo integration to avoid the direct O(n^2) computation.
- Optimization and normalization: The model is trained end-to-end with iterative optimization of disentanglement, MINE, and reconstruction components, while ring-style normalization addresses irregular feature distributions across heterogeneous domains.The implementation uses feature generators such as LeNet, AlexNet, or ResNet and follows the alternating procedure in Algorithm 1.
4. Experiments
Experiments evaluate DADA for domain-agnostic transfer across digit, Office-Caltech10, and DomainNet settings, including ablations, feature visualizations, convergence, and alignment analyses. Across these evaluations, DADA achieves strong accuracy and produces more class-separated, lower-discrepancy features than comparison methods.
- Digit recognition: 62.3% average accuracy is achieved by model IV on Digit-Five, significantly outperforming other baselines on most domain-agnostic tasks.Model IV incrementally includes class disentanglement, domain disentanglement, ring loss, and reconstruction loss.
- Digit recognition: DADA features are more separated between classes than UFDN and MCD features on the agnostic target domain in the sv→mm,mt,up,sy task.The comparison uses t-SNE embeddings of source-only, UFDN, MCD, and DADA representations.
- Office-Caltech10: On W→A,C,D, DADA reaches 93.1% accuracy versus MCD’s 84.3% and distinguishes class pairs that MCD mainly confuses.The cited confusion matrices compare calculator–keyboard, backpack–headphones, and monitor–projector errors.
- Analysis: DADA features have smaller A-distance than raw ResNet and MCD features, while training losses converge and accuracy becomes steady after about 20 epochs.A-distance is evaluated on W→A,C,D and D→A,C,W; convergence is analyzed on C→A,D,W.
5. Conclusion
The paper introduces DAL and DADA to learn disentangled domain-invariant features for unlabeled, heterogeneous target domains. It reports that class disentanglement, mutual-information minimization, and ring-loss normalization contribute to DADA's empirical efficacy.
- DADA disentangles domain-invariant features in latent space for domain-agnostic learning.
- Class disentanglement and mutual-information minimization enhance feature disentanglement.
- Ring-loss-style normalization boosts DADA's performance on the DAL task.
- Extensive evaluation on DAL benchmarks demonstrates efficacy against several state-of-the-art domain-adaptation algorithms.
A. Model Architecture
The supplementary material provides detailed model architectures for DADA's components and for the Digit-Five, Office-Caltech10, and DomainNet settings.
- Detailed architectures are provided for the Generator, Disentangler, Domain Classifier, Classifier, and MINE components.
- The Digit-Five architecture lists convolutional and fully connected layer dimensions, kernel settings, strides, padding, and dropout probabilities.
B. Details of datasets
The supplementary material documents dataset information, including domain splits for Digit-Five and DomainNet and image counts for Office-Caltech10.
- Dataset details are provided in Table 6.
- For Digit-Five and DomainNet, the material gives train/test splits for each domain.
- For Office-Caltech10, the material gives the number of images in each domain.
- The Office-Caltech10 and DomainNet model architectures are provided separately in Table 7.