PaperlayerView exact revision on arXiv

Source-linked AI summary

State-Relabeling Adversarial Active Learning

Beichen Zhang, Liang Li, Shijie Yang, Shuhui Wang, Zheng-Jun Zha, Qingming Huang

arXiv:2004.04943v1cs.CVcs.LG

TL;DR

Active learning seeks to reduce labeling costs by selecting the most informative unlabeled samples, but selection is challenged by limited annotation and state information. SRAAL combines annotation-embedded representations with an adversarial state discriminator using online uncertainty relabeling, and adds k-center initialization. Experiments on classification and segmentation show that SRAAL outperforms previous state-of-the-art methods, while the initialization algorithm improves performance.

  • Problem

    Active learning must select informative samples with limited labeled data, while early small labeled pools restrict supervised selection and initialization is often random.

  • Method

    SRAAL combines reconstruction and annotation information in a unified representation, relabels unlabeled states using online uncertainty, and initializes the labeled pool with k-center sampling.

  • Results

    Experiments on image classification and segmentation show that SRAAL outperforms previous state-of-the-art active learning methods, while its initial sampling algorithm improves performance.

  • Takeaways & Limitations

    Using annotation and state information together, plus diverse initial coverage, supports more informative sampling in the evaluated active-learning tasks.

  • Takeaways & Limitations

    The formal setup assumes a large unlabeled pool and obtains annotations for an initially selected set through an oracle.

Abstract

from arXiv · show

Active learning is to design label-efficient algorithms by sampling the most representative samples to be labeled by an oracle. In this paper, we propose a state relabeling adversarial active learning model (SRAAL), that leverages both the annotation and the labeled/unlabeled state information for deriving the most informative unlabeled samples. The SRAAL consists of a representation generator and a state discriminator. The generator uses the complementary annotation information with traditional reconstruction information to generate the unified representation of samples, which embeds the semantic into the whole data representation. Then, we design an online uncertainty indicator in the discriminator, which endues unlabeled samples with different importance. As a result, we can select the most informative samples based on the discriminator's predicted state. We also design an algorithm to initialize the labeled pool, which makes subsequent sampling more efficient. The experiments conducted on various datasets show that our model outperforms the previous state-of-art active learning methods and our initially sampling algorithm achieves better performance.

1. Introduction

Active learning reduces labeling demands by selecting informative unlabeled samples, but small early labeled pools and random initialization limit selection quality. SRAAL combines annotation and state information with state relabeling, unified representations, and k-center initialization, and outperforms prior methods on classification and segmentation tasks.

  • 1. Introduction: Early active-learning iterations are difficult because the labeled pool is small, limiting supervised models’ ability to identify highly informative samples.Existing approaches commonly rely on labeled annotations and may require a certain amount of labeled data for high accuracy.
  • 1. Introduction: SRAAL combines annotation and labeled/unabeled state information to select more informative unlabeled samples.Its discriminator uses online uncertainty-based state relabeling to assign different importance to unlabeled data.
  • 1. Introduction: The proposed k-center initialization selects a diverse labeled-pool cover under minimax distance, making subsequent sampling more efficient.The method addresses the common use of random initialization in prior active-learning approaches.
  • 1. Introduction: Experiments on four datasets across image classification and segmentation show that SRAAL outperforms previous state-of-the-art active learning methods.The study also evaluates the online uncertainty indicator, supervised target learner, and initial sampling algorithm through ablations.
  • 1. Introduction: SRAAL generates unified image representations by combining unsupervised reconstruction with supervised annotation information.The reconstruction module and supervised target learner produce complementary representations that embed annotation information.

2. Related work

Related active-learning methods mainly use pool-based distribution or uncertainty criteria, while synthesizing approaches generate informative samples with generative models. Existing approaches face scalability, computational-efficiency, instability, or reliability limitations, motivating alternative selection strategies.

  • 2. Related work: Synthesizing approaches generate informative samples with GANs or VAEs instead of querying an unlabeled pool.The passage identifies high computational complexity and instability as disadvantages of these methods.
  • 2. Related work: Pool-based active learning selects existing unlabeled samples using distribution-based diversity or uncertainty-based criteria.Distribution methods seek diverse labeled pools, whereas uncertainty methods sample points estimated to be uncertain.
  • 2. Related work: Traditional uncertainty methods include Bayesian models, Monte Carlo Dropout, ensembles, and Learning-Loss networks.The cited approaches estimate uncertainty through probabilistic models, multiple forward passes, ensembles, or predicted task loss.
  • 2. Related work: Several established approaches do not scale efficiently to deep networks and large datasets, while combined-method results may be unreliable.The cited limitations include computational inefficiency for ensembles and dropout and unreliability for a combined approach.
  • 2. Related work: Some methods combine distribution and uncertainty in two-step procedures to obtain more generalized or accurate uncertainty estimates.Adversarial approaches such as DFAL and VAAL are described among methods combining these ideas.

3. Method

SRAAL combines annotation-aware representation learning with state-based adversarial sampling to identify informative unlabeled samples. It also uses uncertainty-based state relabeling and a distance-based initialization procedure to improve pool construction and selection.

  • 3.2. Unified representation generator: SRAAL combines a unified representation generator with a labeled/unlabeled state discriminator to select informative samples.The generator learns annotation-embedded image features, while the discriminator predicts state values for sampling.
  • 3.2. Unified representation generator: The generator concatenates unsupervised reconstruction and supervised target-learning representations, using unlabeled and labeled data respectively.UIR reconstructs both labeled and unlabeled images, whereas STL predicts annotations using only labeled samples.
  • 3.3. State discriminator and state relabeling: The online uncertainty indicator relabels unlabeled states according to prediction uncertainty, with scores decreasing as maximum probability increases and increasing with distribution concentration.The score is bounded in [0,1) and is computed from the target model’s prediction vector.
  • 3.4. Sampling strategy: At each iteration, the generator produces representations, the discriminator predicts state values, and the top-K unlabeled samples are sent to the oracle.The discriminator is trained using labeled and relabeled-unlabeled state information.
  • 3.5. Initially sampling algorithm: The initialization algorithm first trains an unsupervised reconstructor, then greedily selects a k-center subset using Euclidean distances between latent variables.This procedure treats initial sampling as a set-cover problem and replaces the usual random initialization.

4. Experiment

The experiments evaluate SRAAL for image classification and compare its initialization strategy with random initialization. Image-classification results are reported on CIFAR-10 and CIFAR-100, with repeated trials using controlled labeled pools.

  • Experimental setup: The evaluation covers image classification and segmentation, with iterative annotation until 40% of the training data is labeled.Experiments begin with 10% labeled data, add 5% per iteration, and average results over five initializations.
  • Image classification: Figure 3 reports active-learning image-classification results on CIFAR-10 and CIFAR-100.The study also includes a comparison between the proposed initialization and random initialization, plus an ablation study.

4.1. Active learning for image classification

Across CIFAR-10, CIFAR-100, and Caltech-101, SRAAL generally outperforms prior active-learning methods, including under limited labeled-data conditions.

  • On CIFAR-10, SRAAL reaches 92.48% accuracy using 40% of the samples, versus 93.5% for ResNet-18 trained on the full dataset.The gap is 1.02 percentage points.
  • On CIFAR-10, SRAAL outperforms MC-Dropout, random sampling, core-set, and VAAL, while alternating with LL4AL across sampling percentages.LL4AL leads at 25%, 35%, and 40%, while SRAAL leads at 15%, 20%, and 30%.
  • SRAAL outperforms previous methods across CIFAR-10, CIFAR-100, and Caltech-101 image-classification experiments.The supplied results report especially strong performance on CIFAR-100 and Caltech-101.
  • The results associate SRAAL’s performance with combining annotation information, labeled/unlabeled state information, state relabeling, and annotation-embedded representations.The supplied discussion specifically contrasts SRAAL with methods using only annotation information.
  • On Caltech-101, SRAAL outperforms all previous methods from the first through the final iteration, with a larger gap over the second-best method than on CIFAR-100.The result is reported for a dataset with fewer and unevenly distributed images per class.

4.2. Active learning for semantic segmentation

On Cityscapes semantic segmentation, SRAAL performs better than the compared active-learning methods, including VAAL, on the challenging segmentation task.

  • The semantic-segmentation evaluation uses 3,475 Cityscapes frames converted into 19 classes, with the same initial labeled pool and selection budget for every method.Semantic segmentation is presented as a more challenging task than image classification.
  • SRAAL outperforms VAAL by a large margin on Cityscapes semantic segmentation.The comparison uses DRN as the target model and mean IoU as the performance measure.
  • SRAAL and VAAL outperform SA, QBC, and MC-Dropout on the Cityscapes experiment.Both methods introduce labeled/unlabeled state information into sample selection.
  • The reported advantage over VAAL is attributed to SRAAL’s online uncertainty indicator and state relabeling for guiding the discriminator toward informative samples.

4.3. Initialization algorithm comparison

SRAAL’s k-center initialization produces a more informative initial labeled pool than random initialization, improving accuracy and reducing variability.

  • The k-center initialization achieves significantly higher mean accuracy than random initialization on CIFAR-10.The evaluated active-learning model is SRAAL with a ResNet-18 target classifier.
  • The k-center initialization has lower standard deviation than random initialization.The figure’s dotted lines represent standard deviation.
  • The authors conclude that the initialization algorithm makes subsequent sampling more efficient.

4.4. Ablation study

The complete SRAAL consistently outperforms ablated variants, while removing both tested modules produces the lowest ablation accuracy.

  • Complete SRAAL consistently outperforms all ablations on the CIFAR-100 active-learning experiment.
  • The ablation removing both the online uncertainty indicator and supervised target learner achieves the lowest accuracy among the ablations.
  • Removing either state relabeling or the annotation-embedded unified representation reduces active-learning performance relative to complete SRAAL.

4.5. Comparison on different uncertainty estimators

The proposed uncertainty indicator is compared with entropy and standard deviation under different sampling ratios, and it outperforms these alternatives in the reported experiment.

  • 4.5. Comparison on different uncertainty estimators: The experiment replaces the proposed uncertainty score with common uncertainty acquisition functions to assess its suitability for state relabeling.The compared alternatives are entropy and standard deviation.
  • 4.5. Comparison on different uncertainty estimators: The proposed uncertainty indicator outperforms entropy and standard deviation under different sampling ratios.The comparison evaluates whether the score better reflects the importance of unlabeled data for state relabeling.

5. Conclusion

The paper proposes SRAAL, which combines annotation and labeled/unlabeled state information to select informative samples, and reports improvements over prior active learning methods and with its initialization strategy.

  • 5. Conclusion: SRAAL combines annotation-embedded image features with an online-updated labeled/unlabeled state discriminator to select informative samples.Its representation generator integrates annotation information with image reconstruction, while the discriminator uses the online indicator.
  • 5. Conclusion: The k-center approach initializes the labeled pool to make subsequent sampling more efficient.The initialization algorithm is presented as an additional component of the method.
  • 5. Conclusion: Experiments on image classification and segmentation show that SRAAL outperforms previous state-of-the-art methods.The conclusion also reports that the initial sampling algorithm significantly improves the model's performance.
Loading 2004.04943v1…