Online Variance Reduction for Domain Adaptation on Streaming Data

2026-07-22 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

The ARROW (Adaptive vaRiance Reduction via Online reWeighting) algorithm is introduced as the first online stochastic variance reduction (SVR) method specifically designed for Maximum Mean Discrepancy (MMD) and Correlation Alignment (CORAL) loss functions on streaming data. This paper addresses the incompatibility of existing offline SVR algorithms with online, distributed, or incremental learning environments. ARROW operates by maintaining moving average references of alignment statistics and adaptively reweighting incoming minibatches to align them with these references. A relaxed reweighting scheme is proposed to ensure tractable weight optimization. Experimental results demonstrate that ARROW achieves competitive performance compared to offline algorithms across metrics such as runtime, the degree of variance reduction, and target domain accuracy.

Key takeaway

For Machine Learning Engineers developing systems with streaming data and domain shift, ARROW offers a robust solution for online domain adaptation. You should consider integrating ARROW to maintain model performance without requiring full dataset retraining, especially when using MMD or CORAL loss functions. This approach allows for continuous learning and adaptation in dynamic environments, ensuring your models remain accurate and efficient.

Key insights

ARROW is the first online SVR algorithm for MMD and CORAL, enabling domain adaptation on streaming data.

Principles

• Online SVR is crucial for streaming data domain adaptation.
• Adaptive reweighting can align minibatch and reference statistics.
• Relaxed reweighting ensures optimization tractability.

Method

ARROW maintains moving average references of alignment statistics and adaptively reweights incoming minibatches to align them with these references, using a relaxed reweighting scheme for tractable optimization.

In practice

• Apply ARROW for real-time domain adaptation.
• Integrate ARROW into distributed learning systems.
• Utilize ARROW for incremental model updates.

Topics

• Online Learning
• Domain Adaptation
• Stochastic Variance Reduction
• MMD Loss
• CORAL Loss
• Streaming Data

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Variance-reduced Domain Adaptation using Paired Sampling — PSDA reduces variance in UDA losses by pairing observations, improving target domain accuracy.
• Meta-Learning Transformers to Improve In-Context Generalization — Training transformers on diverse small datasets improves in-context generalization and modularity over large, uncurated corpora.
• Adaptive Bayesian Online Learning via Expert Aggregation — Aggregating Bayesian update rules as experts enables adaptive, uncertainty-aware prediction on data streams without fixed inferential choices.
• AdaGrad-Diff: A New Version of the Adaptive Gradient Algorithm — AdaGrad-Diff improves optimization robustness by adapting stepsizes based on gradient differences, not just gradient magnitudes.
• FedDAF: Federated Domain Adaptation Using Model Functional Distance — FedDAF improves federated domain adaptation by aggregating models based on functional distance from mean gradient fields, addressing data scarcity and domain shift.
• Parameter-Efficient Vision-Language Adaptation with Continuous Metadata Conditioning for Animal Re-Identification — Continuous metadata conditioning enhances vision-language models for robust animal ReID without test-time metadata.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.