Source-linked AI summary
Discriminative Active Learning
Daniel Gissin, Shai Shalev-Shwartz
TL;DR
Expensive labeling motivates active learning methods that select the most valuable samples under a fixed budget, especially for neural networks and large query batches. DAL formulates batch selection as binary discrimination between labeled and unlabeled examples using learned representations. On image-classification tasks, DAL performs on par with state-of-the-art methods at medium and large batch sizes, while the experiments find no clear advantage over uncertainty sampling at relatively large batch sizes.
Problem
Active learning seeks to reduce expensive labeling by selecting samples that maximize test-accuracy improvement under a fixed budget, but large batches and task-specific methods create practical challenges.
Method
DAL poses active learning as binary classification between labeled and unlabeled examples, selecting queries using a learned representation rather than a trained task classifier.
Results
DAL performs on par with state-of-the-art image-classification methods for medium and large query batch sizes, while large-batch baselines are generally within the statistical margin of error.
Takeaways & Limitations
DAL is a simple method that can conceptually extend beyond classification, while relatively large batches do not clearly favor current state-of-the-art methods over uncertainty sampling.
Takeaways & Limitations
Using unlabeled-pool data to learn the representation could benefit DAL but is left as future work.
Abstract
from arXiv · showhide
We propose a new batch mode active learning algorithm designed for neural networks and large query batch sizes. The method, Discriminative Active Learning (DAL), poses active learning as a binary classification task, attempting to choose examples to label in such a way as to make the labeled set and the unlabeled pool indistinguishable. Experimenting on image classification tasks, we empirically show our method to be on par with state of the art methods in medium and large query batch sizes, while being simple to implement and also extend to other domains besides classification tasks. Our experiments also show that none of the state of the art methods of today are clearly better than uncertainty sampling when the batch size is relatively large, negating some of the reported results in the recent literature.
1 Introduction
Active learning reduces costly labeling by selecting informative samples under a fixed budget. DAL addresses large-batch and task-specific limitations by framing selection as binary classification using learned representations.
- Active learning setting: Active learning selects samples expected to improve test accuracy under a fixed labeling budget.Pool-based sampling iteratively chooses examples from a large unlabeled dataset U, obtains oracle labels, and adds them to labeled set L.
- Limits of prior methods: Traditional heuristic methods often select one sample at a time using a model trained on the labeled dataset.This approach is impractical for large-scale collection and primarily targets classification tasks.
- Task generality: Prior active learning solutions were mostly classification-specific, leaving adaptation to other domains as a separate challenge.The second problem had not been explicitly addressed according to the authors.
- Batch-mode challenge: Batch-mode active learning queries multiple examples per iteration but must avoid selecting highly similar samples that can underperform random sampling.Diversity is therefore important when choosing the top-K examples.
- DAL contribution: DAL uses a learned representation Ψ(x), poses selection as binary classification, and compares favorably with state-of-the-art image-classification methods.It avoids assuming a classifier trained on the labeled set when choosing the query batch.
2 Related work
Related active learning methods use uncertainty, margins, expected model change, or core-set coverage to choose informative or diverse samples. These approaches differ in their model requirements, implementation complexity, and scalability.
- Uncertainty-based methods: Uncertainty-based methods select examples using model output confidence and have performed well across diverse datasets with simple implementations.Neural-network softmax scores can be unreliable because classification networks may be over-confident.
- Margin-based methods: Margin-based methods choose examples near a model’s decision boundary, although their ranking can resemble uncertainty sampling for neural networks.DAL’s empirical analysis reports that the margin-based method ranks unlabeled examples similarly to uncertainty sampling.
- Expected model change: Expected model change methods select examples predicted to produce large parameter updates, often using expected gradient length.The expectation is taken over posterior label distributions induced by the trained classifier.
- Core-set selection: Core-set methods formulate batch selection as robust ε-covering of the unlabeled set in learned-representation space to obtain diverse queries.Their robust integer-program formulation is less straightforward and less scalable to large unlabeled datasets than DAL’s method.
3 DAL - Discriminative Active Learning
DAL selects unlabeled examples by framing active learning as binary discrimination between labeled and unlabeled data, using a learned representation rather than a task-specific classifier. Its domain-adaptation view motivates greedily reducing distributional differences, while related comparisons clarify its density-sensitive sampling behavior and future extensions.
- Motivation: DAL motivates labeling examples so the labeled set represents the unlabeled pool’s distribution across high-dimensional, multi-modal data.This avoids explicitly modeling the pool with density-based or parameterized generative methods.
- DAL formulation: The method casts active learning as binary classification that distinguishes labeled examples L from unlabeled examples U.It operates on a learned representation Ψ(x), with binary labels indicating membership in L or U.
- DAL formulation: At each iteration, DAL minimizes binary log loss over U ∪ L, then selects the top-K unlabeled examples according to the discriminator’s output.The discriminative component does not use labels directly, but the representation Ψ(x) is learned using labeled data; using the original input space hurts performance in practice.
- Relation to domain adaptation: DAL is a greedy proxy for minimizing the H-divergence between labeled and unlabeled distributions by moving selected examples from U to L.This connects the method to domain adaptation, where smaller H-divergence supports smaller differences between source- and target-domain performance.
- Extensions and related methods: DAL decouples representation learning from sample selection and leaves joint domain-adaptation training with unlabeled data as future work.The method is also related to GAN objectives, but changes the labeled distribution only through discrete transfers from U to L rather than differentiable generator updates.
- Difference from the core set approach: Unlike core-set selection, DAL aims to preserve the labeled-to-unlabeled ratio across the manifold, sampling in proportion to density rather than overrepresenting sparse regions.Synthetic experiments confirm this intuition, although sparse-region bias can still be advantageous in some situations.
4 Practical considerations
DAL’s greedy top-K selection can be adapted to large batches by using sequential mini-queries, while its discriminator is trained to fit membership distinctions rather than generalize. The implementation therefore trades runtime against batch-diversity awareness through the number of mini-queries.
- Batch construction: DAL splits a batch of size K into n mini-queries, moving selected examples from U to L before labeling and retraining the binary classifier.This sequential updating makes later selections less likely to duplicate earlier examples, improving batch diversity without changing the fixed representation.
- Batch construction: The parameter n controls the trade-off between DAL’s running time and its awareness of batch diversity.More repeated mini-query steps provide more opportunities to account for previously selected examples.
- Binary classifier: DAL uses a three-hidden-layer multilayer perceptron with width 256 and ReLU activations for the binary classification task.The architecture was chosen because the learned representation was not expected to give a convolutional architecture an advantage.
- Binary classifier: The binary classifier is trained to fit the membership training set and detect examples that are harder to fit, rather than to generalize.A small network or early stopping can help prevent the classifier from fitting the training set perfectly.
5 Empirical evaluation
The evaluation compares DAL with established active-learning baselines across datasets and query batch sizes. Results show that batch size strongly affects relative performance, while ranking comparisons indicate DAL selects examples differently from the baselines.
- Experimental setup: Experiments evaluate MNIST and CIFAR-10 using query batches spanning tens to thousands of examples.The study uses LeNet for MNIST and VGG-16 for CIFAR-10, with results averaged across repeated experiments and reported with empirical standard deviations.
- Evaluated algorithms: The comparison includes Random, Uncertainty, DBAL, DFAL, EGL, Core-Set, and DAL.Uncertainty reports the best of max-entropy and variation-ratio acquisition functions; DBAL adds Monte Carlo dropout confidence estimates.
- Results: Aside from Core-Set at very small batch sizes and EGL generally, all algorithms consistently outperform random sampling.The authors attribute discrepancies involving prior EGL results partly to differences between recurrent and convolutional architectures, while identifying an implementation error for one prior uncertainty result.
- Results: At relatively large batch sizes, baseline methods perform within the statistical margin of error, whereas uncertainty methods lead at batch sizes of 10 and 100.For CIFAR-10 batches in the thousands, all methods have essentially the same performance while clearly beating random sampling.
- Ranking comparison: Uncertainty, DBAL, and DFAL produce similar rankings of the MNIST unlabeled set, but DAL's ranking is completely uncorrelated with theirs.Core-Set is excluded from ranking comparisons because it outputs only a query batch, not a ranking over the unlabeled set.
6 Conclusion
DAL reframes active learning as binary classification to make labeled and unlabeled data indistinguishable, achieving state-of-the-art-parallel results for medium and large image-classification batches.
- DAL poses active learning as binary classification, selecting examples to make the labeled set indistinguishable from the unlabeled pool.
- DAL performs on par with state-of-the-art active learning methods for image classification at medium and large query batch sizes.
- DAL is simple to implement and can conceptually extend to domains beyond classification.
- Uncertainty- and margin-based methods show strong similarity, whereas DAL is completely different while obtaining comparable results.