Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Bayesian Domain Adaptation with Gaussian Mixture Domain-Indexing
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 GMDI, a Bayesian domain adaptation method that models global domain indices as a Gaussian mixture with CRP-determined (stick-breaking, truncated) number of components, improving on VDI for classification and regression.
Problem
Domain adaptation when domain indices are unavailable: existing variational domain index methods assume a single Gaussian prior over global domain indices, which limits adaptation to diverse target domains.
Approach
Replace the single Gaussian prior on global domain indices with a dynamic Gaussian mixture whose number of components is set by a Chinese Restaurant Process (finite stick-breaking truncation), trained via a structured variational ELBO plus an adversarial domain-identity discriminator.
Key results
- On classification, GMDI reaches 99.3% accuracy on DG-60 and 96.9% on Circle, up to 3.4% above VDI.
- On TPT-48 N(24)->S(24), GMDI reduces average MSE from 3.160 (VDI) to 2.493, a 21% reduction.
- On TPT-48 W(6)->E(42), GMDI reduces MSE from 2.496 to 2.087 (about 16%); CompCars accuracy rises to 44.4% vs 42.5% for VDI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| model global domain indices as a dynamic Gaussian mixture rather than a single Gaussian | VDI · Circle | accuracy: 94.3 -> 96.9 | Helped |
| model global domain indices as a dynamic Gaussian mixture rather than a single Gaussian | VDI · DG-15 | accuracy: 94.7 -> 96.5 | Helped |
| model global domain indices as a dynamic Gaussian mixture rather than a single Gaussian | VDI · DG-60 | accuracy: 95.9 -> 99.3 | Helped |
| model global domain indices as a dynamic Gaussian mixture rather than a single Gaussian | VDI · CompCars | accuracy: 42.52 -> 44.4 | Helped |
| model global domain indices as a dynamic Gaussian mixture rather than a single Gaussian | VDI · TPT-48 W (6) → E (42), average of all 42 domains | MSE: 2.496 -> 2.087 | Helped |
| model global domain indices as a dynamic Gaussian mixture rather than a single Gaussian | VDI · TPT-48 N (24) → S (24), average of all 24 domains | MSE: 3.160 -> 2.493 | Helped |
| add CRP-based adaptive mixture-component selection instead of fixed components with Gumbel-Softmax (ablation) | GMDI · TPT-48 W → E | MSE: 2.471 -> 2.087 | Helped |
| add CRP-based adaptive mixture-component selection instead of fixed components with Gumbel-Softmax (ablation) | GMDI · TPT-48 N → S | MSE: 3.050 -> 2.493 | Helped |
| add CRP-based adaptive mixture-component selection instead of fixed components with Gumbel-Softmax (ablation) | GMDI · CompCars | accuracy: 43.0 -> 44.4 | Helped |
Surprising findings
- Several baselines (DANN, ADDA, CDANN, MDD, SENTRY) perform near random guess on Circle, DG-15 and DG-60, and most adaptation methods underperform Source-only on TPT-48 regression.
Limitations
- Still relies on availability of domain identities and cannot infer them as latent variables.
- Number of mixture components K must be chosen as an upper bound; best K differs by dataset (2 for most, 3 for CompCars).
Experimental setup
Datasets: Circle DG-15 DG-60 TPT-48 CompCars
Benchmarks: Circle DG-15 DG-60 TPT-48 CompCars
Metrics: accuracy mean squared error
Baselines: Source-only DANN ADDA CDANN MDD SENTRY D2V VDI
Builds on: VDI (Variational Domain Index) DANN Chinese Restaurant Process stick-breaking construction
Closest papers by meaning
- Domain-Indexing Variational Bayes: Interpretable Domain Index for Domain Adaptation (ICLR 2023 · similarity 0.89)
- Variational Model Perturbation for Source-Free Domain Adaptation (NeurIPS 2022 · similarity 0.87)
- Domain Adaptation as a Problem of Inference on Graphical Models (NeurIPS 2020 · similarity 0.87)
- Unsupervised Visual Domain Adaptation: A Deep Max-Margin Gaussian Process Approach (CVPR 2019 · similarity 0.87)
- A General Representation-Based Approach to Multi-Source Domain Adaptation (ICML 2025 · similarity 0.85)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.