Solving the Cold Start Problem in Active Learning

1,286 concepts877 papers7 proposals

Active learning operates on a compelling premise: a machine learning model can achieve higher accuracy with fewer labeled training instances if it is allowed to choose the data from which it learns. Typically, a model evaluates a pool of unlabeled data, identifies the samples about which it is most uncertain, and requests labels for those specific points from a human expert.

However, this framework suffers from a fundamental circular dependency known as the cold-start problem. To select the most informative samples, the active learning algorithm relies on the model's internal confidence or uncertainty metrics. But at the very beginning of training—when zero or very few labels are available—the model's parameters are unoptimized. As a result, its predictions are stochastic, its decision boundaries are arbitrary, and its uncertainty estimates are completely uncalibrated.

Relying on an untrained model to guide data collection creates a destructive feedback loop. As noted in the introductory overview Introduction to Active Learning and Application to Material Engineering, an active learner in its early stages often performs worse than simple random sampling. Because the initial model lacks a baseline understanding of the domain, it generates biased queries, focusing on noisy outliers or redundant regions of the feature space while remaining completely blind to entire categories of data.

Why Early Uncertainty Fails

To understand why traditional active learning breaks down at initialization, consider how uncertainty is measured. In classification or sequence labeling tasks, a model might use entropy or margin sampling to find data points near its current decision boundary.

As Vincent Claveau demonstrates, uncertainty-based strategies regularly underperform in these early phases because the model lacks the structural context of the feature distribution. If the initial seed set is not perfectly representative, the model becomes overconfident in incorrect decision boundaries. Anqi Liu et al. formalize this as an issue of "optimism bias": the active learner assumes its current, biased model is a reliable guide for exploration, causing it to get stuck in suboptimal regions of the feature space.

This warm-start dependency means that practitioners traditionally have to collect a substantial, randomly sampled "seed dataset" to bootstrap the model before active learning can even begin. This requirement defeats the purpose of active learning in low-resource domains, where manual annotation is highly expensive or unavailable from the start.

Decoupling Selection from the Predictor

The core insight behind modern solutions to the cold-start problem is decoupling: separating the sample selection mechanism from the task-specific predictor's performance during the initial stages of training. If the selection mechanism does not rely on the model's uncalibrated uncertainty, it cannot be misled by it.

One way to achieve this is by exploiting the fixed structural and geometric properties of the unlabeled data pool itself. For example, in graph-structured data, Yichong Xu et al. introduced a method called FeatProp. Instead of querying nodes based on a Graph Neural Network's fluctuating predictions, the authors propagate and smooth node features across the graph topology, then apply a clustering algorithm to select representative exemplars. Because this selection relies entirely on the static geometry of the feature space, it remains completely stable even when zero labels are initially available.

Another approach is to use an independent model whose sole job is to measure distribution coverage rather than task performance. Yue Deng et al. proposed an adversarial active learning framework (ALISE) that uses an independent discriminator to identify unlabeled samples that look the least like the current training set. By focusing strictly on how much a sample differs from the current training distribution—rather than the predictor's internal confidence—the system bypasses the noise and initialization bias inherent in poorly trained models.

Bootstrapping with Self-Supervision and Foundation Priors

A parallel line of research leverages the power of self-supervised learning and pre-trained foundation models to inject a "pseudo-supervised" signal into the cold-start phase, bypassing the need for initial human labels.

Instead of starting with a blank slate, we can use self-supervised pretext tasks—such as predicting image rotation or data masking—to evaluate how well a model understands the underlying data distribution. John Seon Keun Yi et al. developed PT4AL, a framework based on the discovery that a model's loss on a self-supervised pretext task is highly correlated with its potential loss on the actual downstream task. By selecting the highest-loss samples from the pretext task to form the initial training batch, the authors ensure that the first labeled set contains the most structurally informative examples, significantly outperforming random initialization.

In the natural language processing domain, where large language models possess massive latent knowledge, Yue Yu et al. proposed a prompt-based uncertainty propagation approach. Instead of requiring a labeled warm-start dataset to estimate uncertainty, the authors use carefully designed prompts to elicit the model's own implicit confidence about unlabeled data points. This effectively simulates a supervised signal at zero-shot initialization, allowing the active learning pipeline to select highly beneficial samples before a single human label has been collected.

By shifting the initialization strategy from random sampling to structure-aware or self-supervised selection, these methods successfully break the active learning Catch-22, enabling data-efficient learning from day one.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_active_learning_cold_start_3f34f657,
  title = {active learning cold start},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/active-learning-cold-start/rdr_3f34f657698040a19c2cee87a5c002cb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_active_learning_cold_start_3f34f657_v1,
  title = {active learning cold start},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/active-learning-cold-start/rdr_3f34f657698040a19c2cee87a5c002cb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake