Research map / Continual learning / Domain and prompt-based
CONTRAST: Continual Multi-source Adaptation to Dynamic Distributions
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 CONTRAST, a test-time method that learns convex combination weights over multiple frozen-then-combined source models from streaming small test batches (no source data), and adapts only the most-correlated source with an existing TTA method (Tent, CoTTA, EATA).
Problem
Adapting an ensemble of pre-trained source models to dynamically shifting test distributions that arrive in small streaming batches without access to source data, while avoiding forgetting and matching the best single source.
Approach
Combine source models with weights optimized by entropy minimization over weighted pseudo-labels, initialize weights from BN-statistic KL divergence and pick a closed-form step size, then update only the source with the largest weight using an off-the-shelf single-source TTA method so less-correlated sources are not forgotten.
Key results
- On CIFAR-100C (15 sequential noise domains), Tent+CONTRAST achieves 37.1% mean error vs 68.2% for Tent-Best, and EATA+CONTRAST reaches 34.2% vs 39.9% for EATA-Best.
- On Office-Home, Tent+CONTRAST averages 37.0% error vs 39.9% for Tent-Best and 41.6% for Source-Best.
- On Digits, Tent+CONTRAST averages 14.3% error vs 17.1% for Tent-Best; CONTRAST shows virtually no source-accuracy forgetting in the reported analysis.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine multiple sources with CONTRAST rather than using the best single source | Tent adaptation · CIFAR-100C (15-domain mean) | error rate (%): 68.2 -> 37.1 | Helped |
| Combine multiple sources with CONTRAST rather than using the best single source | EaTA adaptation · CIFAR-100C (15-domain mean) | error rate (%): 39.9 -> 34.2 | Helped |
| Combine multiple sources with CONTRAST rather than using the best single source | CoTTA adaptation · CIFAR-100C (15-domain mean) | error rate (%): 38.5 -> 34.4 | Helped |
| Combine multiple sources with CONTRAST rather than using the best single source | Tent adaptation · Office-Home (average) | error rate (%): 39.9 -> 37.0 | Helped |
| Combine multiple sources with CONTRAST rather than using the best single source | Tent adaptation · Digits (average) | error rate (%): 17.1 -> 14.3 | Helped |
| Replace BN-statistic KL initialization with prediction-entropy initialization (ablation) | CONTRAST · CIFAR-100C (15-domain mean) | error rate (%): 37.1 -> 53.6 | Hurt |
| Replace proposed adaptive step size with fixed step size 1e-3 (ablation) | CONTRAST with proposed initialization · Office-Home | error rate (%): 37.0 -> 37.9 | Hurt |
| Replace proposed adaptive step size with fixed step size 1e-2 (ablation) | CONTRAST with proposed initialization · Office-Home | error rate (%): 37.0 -> 37.8 | Hurt |
| Replace proposed adaptive step size with fixed step size 1e-1 (ablation) | CONTRAST with proposed initialization · Office-Home | error rate (%): 37.0 -> 37.5 | Hurt |
| Replace proposed adaptive step size with fixed step size 1e0 (ablation) | CONTRAST with proposed initialization · Office-Home | error rate (%): 37.0 -> 37.4 | Hurt |
| Replace proposed adaptive step size with fixed step size 1e1 (ablation) | CONTRAST with proposed initialization · Office-Home | error rate (%): 37.0 -> 39.1 | Hurt |
| Use random rather than proposed initialization, with the table's tested fixed step sizes (ablation) | CONTRAST · Office-Home | error rate (%): Random initialization: 40.7, 40.9, 40.6, 39.6, 41.5; Ours initialization with adaptive step size: 37.0 | Hurt |
Surprising findings
- Uniform ensembling of adapted sources performs much worse than the best source (about 45.5% error on SVHN vs 14.2% for CONTRAST), so the authors exclude it as a baseline.
- Updating only the most correlated source model outperforms updating all models or a subset, even for CoTTA, which is designed to resist forgetting.
Limitations
- Assumes all samples within a test batch come from the same distribution.
- Evaluated mainly on ResNet-18 and corruption/digit/office benchmarks; segmentation results are only in the appendix.
Experimental setup
Datasets: CIFAR-100C CIFAR-10C Office-Home Digits
Benchmarks: CIFAR-100C CIFAR-10C Office-Home Digits
Metrics: error rate (%) source-domain accuracy (forgetting analysis)
Baselines: Tent CoTTA EATA DECISION Model Soups Source Best Source Worst
Builds on: Tent CoTTA EATA Model Soups
Closest papers by meaning
- RDumb: A simple approach that questions our progress in continual test-time adaptation (NeurIPS 2023 · similarity 0.88)
- On Pitfalls of Test-Time Adaptation (ICML 2023 · similarity 0.88)
- Bayesian Weight Enhancement with Steady-State Adaptation for Test-time Adaptation in Dynamic Environments (ICML 2025 · similarity 0.86)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.86)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.85)
- Adaptive Data Optimization: Dynamic Sample Selection with Scaling Laws (ICLR 2025 · 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.