In the traditional active learning paradigm, a model learns like a student taking a long series of pop quizzes. It studies a few examples, identifies what it doesn't know, asks a teacher for labels, and repeats this cycle dozens or hundreds of times. While effective, this iterative "train-select-label" loop is becoming a massive computational bottleneck. As models grow into billions of parameters, retraining them every time you want to pick the next ten images to label is no longer feasible.
Researchers are shifting toward One-Shot Active Learning (also known as single-pass or zero-round selection). The goal is to identify the entire budget of most informative samples in a single step, before the main training process even begins.
The Bottleneck of Iteration
Traditional active learning relies on the model's own uncertainty to pick data. However, as noted by Gabriella Contardo, Ludovic Denoyer, and Thierry Artières in A Meta-Learning Approach to One-Step Active-Learning, many real-world scenarios require labeling to occur in a single batch. Sequential selection forces human experts to wait for model updates, creating a "standby" cost that is often more expensive than the labeling itself.
Furthermore, iterative selection can suffer from "representation bias." If early iterations select poor samples, the model’s uncertainty metrics become skewed, leading to a feedback loop of bad data. One-shot methods aim to bypass this by looking at the data's inherent structure rather than a half-trained model's fluctuating opinions.
Leveraging the Pretrained Feature Space
The rise of foundation models has provided a "cheat code" for one-shot selection. Instead of training a model from scratch to find out what is informative, we can use the fixed feature representations of a self-supervised or pretrained model.
Yichen Xie, Masayoshi Tomizuka, and colleagues demonstrate in Towards General and Efficient Active Learning and Active Finetuning that these pretrained spaces are rich enough to identify informative samples without any task-specific labels. By performing a single pass of inference, their FreeSel method achieves a 530x speedup over iterative methods. The intuition is simple: if a pretrained model sees two images as semantically similar, you probably only need to label one of them.
Balancing Diversity and Difficulty
To select a perfect batch in one go, a method must balance two competing needs:
- Representativeness (Diversity): The batch should cover the entire "shape" of the dataset.
- Informativeness (Difficulty): The batch should focus on the "boundaries" where the model is likely to struggle.
In Boundary Matters: A Bi-Level Active Finetuning Method, Han Lu and Junchi Yan propose BiLAF, which uses a bi-level optimization to pick "core" samples for diversity and "boundary" samples for uncertainty. Similarly, Noga Bar and Raja Giryes in Diverse Subset Selection via Norm-Based Sampling and Orthogonality use the Gram-Schmidt process to ensure that every selected sample is "orthogonal" (mathematically diverse) to the ones already picked.
Other researchers use matrix decomposition to solve this. Changsheng Li and Hongyuan Zha introduced ALFS in Joint Active Learning with Feature Selection via CUR Matrix Decomposition, which treats data selection as a reconstruction problem: find the smallest subset of data that can best "reconstruct" the entire original matrix.
One-Shot Selection Across Domains
The "one-shot" philosophy is being applied to diverse fields where iteration is physically impossible or too slow:
- Genomics: In Efficient Data Selection for Training Genomic Perturbation Models, Johannes F. Lutzeyer and Jun Pang use a method called GRAPHREACH to pick gene experiments upfront. This reduced a five-month wet-lab experimental timeline to just one month.
- Federated Learning: Ying-Peng Tang and Han Yu developed FALE (Efficient Heterogeneity-Aware Federated Active Data Selection), which allows a server to pick the best data from remote clients in a single communication pass, saving massive amounts of bandwidth.
- Large Language Models: For fine-tuning LLMs, Gantavya Bhatt and Yifang Chen showed in An Experimental Design Framework for Label-Efficient Supervised Finetuning of Large Language Models that selecting prompts in one shot using a base model's internal representations can halve the annotation bill while maintaining performance.
The Trade-off: Static vs. Dynamic
The primary challenge for one-shot methods is the lack of feedback. As Haoyu Han notes in Structural Fairness-aware Active Learning for Graph Neural Networks, because the entire budget is selected before training begins, the strategy cannot adjust to specific failure modes that emerge as the model learns.
However, the evidence from researchers like Sheng-Jun Huang in One-shot Active Learning Based on Lewis Weight Sampling for Multiple Deep Models suggests that for modern, large-scale deep learning, the efficiency gains of one-shot selection far outweigh the marginal benefits of iterative refinement. By treating data selection as a structural optimization problem rather than a training loop, we can build high-performing models in a fraction of the time.