Research map / Label-efficient and robust learning / Domain adaptation and generalization

Enhancing Evolving Domain Generalization through Dynamic Latent Representations

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Optimization-based domain generalization · 7 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MISTS, a mutual-information-regularized sequential autoencoder with a domain-adaptive classifier that learns both invariant and dynamic latent features for evolving domain generalization, evaluated on six synthetic and real-world EDG benchmarks.

Problem

Evolving domain generalization: predicting labels on unseen future domains that drift continuously over time, where invariant-feature methods alone are insufficient.

Approach

Disentangle invariant and dynamic latent variables in a sequential autoencoder by adding mutual-information terms (minimize MI between z_c and z_t, maximize MI with input), and predict labels with a domain-adaptive classifier whose weights evolve over domains, backed by a causal model and an ELBO derivation.

Key results

• MISTS averages 69.9% accuracy across seven datasets in the main table, versus 67.0% for LSSAE (the best prior EDG baseline) and 62.3% for ERM.
• On Sine, MISTS reaches 78.2% vs. 68.4% for LSSAE; on RMNIST 47.5% vs. 45.3%.
• Ablation on RMNIST: removing the mutual-information terms drops accuracy from 47.5% to 44.3%; removing the adaptive classifier drops it to 42.1%.

Measured effects reported in the paper

Surprising findings

• Dropping the dynamic latent z_t (keeping only invariant features) gives 45.7% on RMNIST, and dropping z_c gives only 10.2%, showing the invariant-only and dynamic-only variants perform very differently.
• In the appendix's first-target-domain table, LSSAE outperforms MISTS on Circle (91.5 vs 88.7), so MISTS does not win on every dataset.

Limitations

• The relative importance of invariant and dynamic features may vary across datasets and over time, and the method does not adapt this automatically.
• Results on Circle in the first-target-domain table favor LSSAE, so gains are not uniform across datasets.
• Standard deviations overlap on several datasets (e.g., Circle, Elec), so some improvements are small.

Experimental setup

Datasets: Circle Circle-C Sine Sine-C Rotated MNIST Portraits Caltran Elec

Benchmarks: Circle Sine Sine-C Elec RMNIST Portraits Caltran

Metrics: accuracy

Baselines: ERM IRM CORAL Mixup LISA GroupDRO EWC SI DRAIN LSSAE

Builds on: LSSAE IRM variational autoencoder

Closest papers by meaning

• Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains (ICML 2025 · similarity 0.90)
• Weight Diffusion for Future: Learn to Generalize in Non-Stationary Environments (NeurIPS 2024 · similarity 0.89)
• Generalizing across Temporal Domains with Koopman Operators (AAAI 2024 · similarity 0.88)
• Continuous Domain Generalization (NeurIPS 2025 · similarity 0.87)
• A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.87)
• Self-Training with Dynamic Weighting for Robust Gradual Domain Adaptation (NeurIPS 2025 · similarity 0.87)