Non-Iterative Active Learning

49 concepts36 papers

Active learning has traditionally been defined by a loop: select a small batch of data, ask a human to label it, retrain the model, and use the updated model's uncertainties to select the next batch. While this iterative process is mathematically elegant, it introduces severe operational bottlenecks in modern machine learning workflows. Retraining a large language model or a massive vision transformer after every labeling step is computationally prohibitive. Furthermore, keeping human experts on standby while models retrain creates significant operational friction, and frequent synchronization between servers and clients in federated environments strains communication bandwidth.

To bypass these bottlenecks, a growing body of research focuses on one-shot (or single-pass) active learning. The core idea is simple yet ambitious: identify the entire optimal budget of training samples in a single pass before any human labeling or model fine-tuning begins.

But this introduces a fundamental paradox: if active learning relies on a model's feedback to find informative data, how can we make optimal selections without an active feedback loop?


Leveraging Pre-Trained Feature Spaces

The most prominent solution to the one-shot paradox relies on the pre-training-and-fine-tuning paradigm. Modern models rarely start learning from scratch; they begin with rich, general-purpose representations inherited from self-supervised pre-training. One-shot active learning can exploit this fixed feature space to map out the structure of unlabeled data before fine-tuning.

Instead of relying on a changing model to flag uncertain points, algorithms can analyze the static geometry of the pre-trained embeddings. For instance, Yichen Xie et al. demonstrated in Towards Free Data Selection with General-Purpose Models that a single-pass inference through a general-purpose model can identify diverse semantic patterns across an image pool, achieving massive speedups over traditional active learning loops.

By looking at data through the lens of a frozen, robust feature extractor, selection algorithms can balance two critical dimensions of data utility in one go:

  1. Global Diversity (Core Samples): Ensuring the selected batch covers the entire distribution of the dataset.
  2. Local Uncertainty (Boundary Samples): Identifying points that lie near decision boundaries.

In Boundary Matters: A Bi-Level Active Finetuning Method, Han Lu et al. showed that balancing these two dimensions within a pre-trained feature space allows for highly accurate data selection in a single shot. This perspective is echoed by Qian Shao et al. in Enhancing Semi-Supervised Learning via Representative and Diverse Sample Selection, confirming that pre-trained spaces can effectively isolate high-utility samples without the risk of error propagation common in iterative model-in-the-loop selection.


Simulating Feedback with Proxy Models

When a pre-trained model is not perfectly aligned with the target task, researchers use lightweight "surrogates" or proxy models to estimate data utility without triggering full retraining cycles.

One approach is to learn a standalone data utility function from a tiny, initial seed of labeled data. As developed by Jiachen T. Wang et al. in One-Round Active Learning through Data Utility Learning and Proxy Models, this utility function acts as a proxy that scores the expected performance gain of unlabeled batches, allowing practitioners to optimize and select a large batch all at once.

Similarly, when evaluating massive models where forward passes are expensive, small surrogate models can act as stand-ins. In Scaling Up Active Testing to Large Language Models, Gabrielle Berrada et al. utilized small, fixed surrogate models initialized via in-context learning to perform a single forward pass over an unlabeled pool. This strategy successfully estimates model performance and slashes evaluation costs without requiring iterative, gradient-based fine-tuning.


Mathematical and Structural Invariants

Another way to achieve single-shot selection is to focus on properties of the data that remain invariant regardless of the model's training state. This is particularly powerful in structured domains like graph neural networks (GNNs) or physics-informed neural operators.

In graph-structured data, the topology of the network itself reveals which nodes are most influential. Haoyu Han et al. demonstrated in Structural Fairness-aware Active Learning for Graph Neural Networks that an optimal set of training nodes can be identified in a single pre-training phase by balancing graph representativeness with structural fairness. This decouples the selection process entirely from the specific parameters of the GNN, eliminating the need for expensive iterative optimization over large graphs.

For continuous mathematical problems, physical laws can substitute for label feedback. In PICore: Physics-Informed Unsupervised Coreset Selection for Data Efficient Neural Operator Training, Anant Khandelwal et al. showed that training a model briefly on physical constraints allows for a one-time selection pass to identify a representative coreset before running any expensive numerical simulations.

From a statistical perspective, non-adaptive sampling can achieve near-optimal sample complexity by minimizing generalization bounds directly. JH Krijthe et al. proved in Nuclear discrepancy for single-shot batch active learning that optimizing a proxy for the generalization bound (such as nuclear discrepancy) allows for effective single-shot batch selection. In regression tasks, Aditya Parulekar et al. showed in L1 Regression with Lewis Weights Subsampling that sampling data points in a single batch according to their Lewis weights yields high-probability error guarantees that match the performance of complex, adaptive sequential algorithms.


Simulating the Loop Internally

If an algorithm strictly requires sequential feedback to maintain diversity and avoid selecting redundant samples, that feedback loop can be simulated entirely on the machine's side without waiting for human intervention.

One method is to use pseudo-annotators. In Single shot active learning using pseudo annotators, Yazhou Yang et al. proposed a framework where a model updates itself iteratively using synthetic, random labels for each selected sample. This trick forces the selection algorithm to explore the data distribution and pick a diverse batch sequentially, even though no real human feedback has entered the system yet.

Alternatively, the selection process can be framed as a reinforcement learning problem where agents learn to select batches of data simultaneously. As shown by Hanghang Tong et al. in Batch Active Learning with Graph Neural Networks via Multi-Agent Deep Reinforcement Learning, multi-agent value decomposition can be used to select informative and diverse groups of nodes at the same time, bypassing the sequential bottleneck entirely.


Why One-Shot Selection Works

One-shot active learning works because it shifts the burden of data selection from runtime model state to upstream data structure.

Traditional active learning treats the model as a fragile entity that must be carefully guided at every step of its evolution. One-shot active learning recognizes that between pre-trained representations, structural invariants, and statistical proxies, we already possess the geometric tools necessary to identify where the information density of a dataset resides. By consolidating data selection into a single, high-quality pass, these methods eliminate computational overhead and operational friction while delivering performance competitive with traditional iterative pipelines.

Key Foundational Work

Sample Selection for Statistical Parsers: Cognitively Driven Algorithms and Evaluation Measures (2009) — contributed 1 concept elements to this direction.
Active learning for sense annotation (2015) — 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_single_shot_active_learning_ecbb0c3a,
  title = {single-shot active learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-shot-active-learning/rdr_ecbb0c3ad59b4b93855b6900252e3b78},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_single_shot_active_learning_ecbb0c3a_v1,
  title = {single-shot active learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/single-shot-active-learning/rdr_ecbb0c3ad59b4b93855b6900252e3b78/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}

See something wrong? Report a mistake