Learning Transferable Data Selection Policies

22 concepts20 papers1 proposal

When training machine learning models under tight resource constraints, deciding which data points to label, which historical windows to train on, or which clients to sample is a critical bottleneck. Traditionally, fields like active learning and self-training have relied on hand-crafted, static heuristics—such as selecting instances with the highest predictive entropy or lowest model confidence.

However, these rigid rules assume that a single mathematical formula can universally capture data utility. In practice, they tightly couple data selection to a specific model architecture and a specific dataset.

A rapidly growing body of research shifts this paradigm by treating data selection as a meta-learning problem. The core insight is that data informativeness is governed by universal structural patterns—such as how data points cluster in an embedding space, how uncertainty manifests across model layers, or how a model's loss landscape responds to new information. Because these geometric and statistical properties transcend specific class labels or domains, we can learn a generalized selection policy on a set of source tasks and transfer it to entirely unseen target domains without retraining.

Meta-Learning the Topology of Utility

To build a data selection policy that generalizes across domains, we must train it to recognize the underlying relationship between an instance's features, the current model state, and the resulting downstream performance gain. Instead of evaluating data points in isolation, a selection policy can be trained across a diverse distribution of tasks to observe how different subsets of data accelerate model convergence.

[Source Tasks / Domains] ---> [Observe Training Trajectories] ---> [Learn Universal Selection Policy]
                                                                                |
                                                                                v
[Unseen Target Domain]   <---   [Apply Learned Policy]      <--- [Extract Informative Samples]

For example, in Picking Groups Instead of Samples: A Close Look at Static Pool-Based Meta-Active Learning, Ignasi Mas et al. demonstrated that meta-active learning models can learn to select optimal groups of data points simultaneously. By optimization over diverse training tasks, the selector captures domain-agnostic heuristics for diversity and representativeness that generalize to completely unseen categories. Similarly, Gabriella Contardo et al. introduced a framework in A Meta-Learning Approach to One-Step Active-Learning that uses policy gradients to optimize a one-step data acquisition strategy. Their work shows that exposing the model to a high variety of tasks during meta-training allows the selection logic to transcend specific domains.

Rather than learning from scratch via reinforcement learning, which can be computationally expensive, policies can also be trained via imitation learning by mimicking a "greedy" algorithmic expert. In Learning How to Actively Learn: A Deep Imitation Learning Approach, Ming Liu et al. proved that these learned strategies capture fundamental principles of data utility that remain robust when transferred across different languages and tasks. This scalability is further reinforced by Philip Bachman et al. in Learning Algorithms for Active Learning, where selection strategies learned on simple, low-way classification tasks naturally generalized to highly complex, many-class scenarios.

Decoupling Selection via General-Purpose Representations

Why do these learned policies transfer so effectively without collapsing when encountering new distributions? They succeed because they decouple the selection mechanism from the internal, task-specific parameters of the target model, relying instead on general feature representations.

With the advent of foundation models, this decoupling has become even more powerful. In Towards Free Data Selection with General-Purpose Models, Yichen Xie et al. demonstrated that general-purpose pretrained models possess an intrinsic visual understanding capable of identifying informative semantic patterns for downstream tasks (like object detection or segmentation) without any task-specific fine-tuning.

In the text domain, Seungmin Seo et al. developed Active Learning on Pre-trained Language Model with Task-Independent Triplet Loss, leveraging task-independent triplet losses on top of pretrained language models to build a unified selection mechanism that performs consistently across disparate NLP tasks like relation extraction and sentence classification.

This universality is also evident in generative and representation learning settings. In Autoencoding Conditional Neural Processes for Representation Learning, Victor Prokhorov developed a variational framework that learns to identify highly informative "context points" within an image, demonstrating that the learned selection strategy generalizes zero-shot to entirely different data distributions. By focusing on the underlying data manifold rather than class labels, the selection policy treats informativeness as a geometric property.

Cross-Domain Generalization in Practice

The ultimate proof of this paradigm is its success across highly divergent domains, where a policy trained on one type of data successfully guides learning in an entirely different context.

Computer Vision and Video Analysis

In temporal action localization, where labeling video frames is exceptionally costly, Fabian Caba Heilbron et al. showed in What Do I Annotate Next? An Empirical Study of Active Learning for Action Localization that active learning selection functions trained on a source dataset can successfully identify high-utility samples in target datasets containing entirely different action categories. Similarly, Serena Yeung-Levy et al. used a reinforcement learning agent in Learning to Learn from Noisy Web Videos to filter noisy web search results, demonstrating that the logic for identifying high-quality training examples transfers seamlessly to novel, long-tail classes.

Domain Adaptation and Self-Training

When adapting models to new environments, fixed confidence thresholds often fail due to semantic drift. In Learning How to Self-Learn: Enhancing Self-Training Using Neural Reinforcement Learning, Chenhua Chen et al. replaced these thresholds with a deep Q-network that evaluates instance features and model uncertainty, achieving robust self-training gains across diverse languages and domain shifts.

For complex domain adaptation scenarios involving mismatched categories, Didi Zhu et al. utilized Generative Flow Networks in Generalized Universal Domain Adaptation with Generative Flow Networks to build task-agnostic sampling strategies that transfer across domain pairs without retraining. Even when the source and target distributions diverge significantly, Min Xiao et al. proved in Online Active Learning for Cost Sensitive Domain Adaptation that a cost-sensitive active learning framework can intelligently distribute budgets across domains, showing that domain similarity is not a prerequisite for selection utility.

Time Series, Systems, and Materials Science

The principle extends far beyond traditional vision and language modalities:

By shifting the focus from what the data represents semantically to how the data behaves statistically and geometrically, transferable selection policies uncover a unified, cross-domain logic for data utility.

Key Foundational Work

An Approach to Text Corpus Construction which Cuts Annotation Costs and Maintains Reusability of Annotated Data (2007) — contributed 1 concept elements to this direction.
Online Active Learning for Cost Sensitive Domain Adaptation (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_domain_selection_generalization_1052c412,
  title = {cross-domain selection generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-domain-selection-generalization/rdr_1052c4120ab44b4fbb38fadfce045a19},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_domain_selection_generalization_1052c412_v1,
  title = {cross-domain selection generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-domain-selection-generalization/rdr_1052c4120ab44b4fbb38fadfce045a19/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake