PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Deep Active Learning

Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B. Gupta, Xiaojiang Chen, Xin Wang

arXiv:2009.00236v2cs.LGstat.ML

TL;DR

Deep active learning addresses the need to reduce expensive expert annotation while retaining deep learning’s powerful representation capabilities, a gap arising from differences between AL and DL. This survey systematically classifies DeepAL methods, reviews applications, and discusses unresolved problems and future directions; it concludes that the field has practical value but remains in its infancy.

  • Problem

    Deep learning’s dependence on large labeled datasets conflicts with active learning’s limited-label setting, while existing DeepAL research lacks a comprehensive survey.

  • Method

    The survey classifies and reviews DeepAL methods across query strategies, labeled-sample expansion, model generality, stopping strategies, applications, and future directions.

  • Results

    DeepAL research has produced successful applications and combines deep learning’s feature extraction with active learning’s potential to reduce annotation costs.

  • Takeaways & Limitations

    DeepAL has practical application value for tasks where labels are costly or require expertise, but the research remains at an early stage.

  • Takeaways & Limitations

    DeepAL results vary across studies, motivating unified evaluation platforms, stronger random-sampling baselines, and reproducibility practices.

Abstract

from arXiv · show

Active learning (AL) attempts to maximize the performance gain of the model by marking the fewest samples. Deep learning (DL) is greedy for data and requires a large amount of data supply to optimize massive parameters, so that the model learns how to extract high-quality features. In recent years, due to the rapid development of internet technology, we are in an era of information torrents and we have massive amounts of data. In this way, DL has aroused strong interest of researchers and has been rapidly developed. Compared with DL, researchers have relatively low interest in AL. This is mainly because before the rise of DL, traditional machine learning requires relatively few labeled samples. Therefore, early AL is difficult to reflect the value it deserves. Although DL has made breakthroughs in various fields, most of this success is due to the publicity of the large number of existing annotation datasets. However, the acquisition of a large number of high-quality annotated datasets consumes a lot of manpower, which is not allowed in some fields that require high expertise, especially in the fields of speech recognition, information extraction, medical images, etc. Therefore, AL has gradually received due attention. A natural idea is whether AL can be used to reduce the cost of sample annotations, while retaining the powerful learning capabilities of DL. Therefore, deep active learning (DAL) has emerged. Although the related research has been quite abundant, it lacks a comprehensive survey of DAL. This article is to fill this gap, we provide a formal classification method for the existing work, and a comprehensive and systematic overview. In addition, we also analyzed and summarized the development of DAL from the perspective of application. Finally, we discussed the confusion and problems in DAL, and gave some possible development directions for DAL.

1 INTRODUCTION

Deep learning provides powerful feature extraction but depends heavily on labeled data, while active learning seeks useful samples to reduce annotation costs. Deep active learning combines these complementary goals, and this survey organizes its methods, applications, challenges, and future directions.

  • Deep learning automatically extracts features through complex models but requires many labeled samples for training.
  • Active learning selects high-value samples from unlabeled data for oracle labeling to reduce annotation costs while maintaining performance.
  • Deep active learning combines deep learning’s high-dimensional feature extraction with active learning’s potential to reduce labeling costs.
  • The survey presents a comprehensive review, classification, application analysis, and discussion of challenges and future directions for deep active learning.

2 THE NECESSITY AND CHALLENGE OF COMBINING DL AND AL

Combining deep learning and active learning is motivated by their complementary strengths but creates challenges because their data requirements, uncertainty estimates, and processing pipelines differ.

  • Deep active learning is proposed to combine deep learning’s feature extraction and high-dimensional data processing with active learning’s labeling-cost reduction.
  • Deep learning can be overconfident, making softmax response unreliable for uncertainty-based querying and potentially worse than random sampling.
  • Active learning’s small labeled samples and one-by-one queries may be insufficient or unsuitable for data-hungry deep learning.
  • Researchers address these challenges with Bayesian deep learning, data augmentation, pseudo-labeling, and combined supervised–semisupervised training.

3 DEEP ACTIVE LEARNING

A typical DeepAL framework uses a deep model to extract features from an unlabeled pool, queries selected samples, retrains on the expanded labeled set, and repeats the cycle.

  • DeepAL initializes or pre-trains model parameters on an initial labeled set before extracting features from the unlabeled pool.
  • A query strategy selects samples for oracle labeling, after which the new labeled set trains the deep learning model and updates the unlabeled pool.
  • The DeepAL cycle repeats until the label budget is exhausted or predefined stopping conditions are reached.

3.1 Query Strategy Optimization in DeepAL

DeepAL query-strategy research centers on selecting small batches that balance informativeness, diversity, and uncertainty while accounting for deep-learning constraints. The survey organizes these strategies and highlights how batch correlation, dataset characteristics, and scalable querying shape their design.

  • Query-strategy foundations: DeepAL formulates query selection as choosing a small labeled set from an unlabeled pool while maintaining a predetermined accuracy level.The query strategy Q maps the unlabeled dataset to the current labeled set, with the target m ≪ n.
  • Batch Mode DeepAL: Batch-based querying replaces one-by-one selection because frequent retraining changes little data and can be inefficient or encourage overfitting in deep models.The survey identifies batch mode as the foundation of DeepAL query strategies.
  • Uncertainty-based strategies: BALD ranks samples by mutual information between model parameters and predictions, but independent selection can produce information-rich yet highly similar batches.Higher mutual information indicates greater sample uncertainty, while the independent formulation ignores correlations among queried samples.
  • Batch Mode DeepAL: BatchBALD addresses batch redundancy by estimating joint mutual information between multiple queried labels and model parameters.Its formulation explicitly considers correlations among different query samples rather than evaluating each sample independently.
  • Uncertainty-based and Hybrid Query Strategies: Hybrid strategies combine uncertainty or information volume with sample diversity because insufficient diversity can underuse the data distribution and impair deep-model training.Exploration-P seeks high uncertainty and low redundancy, while DMBAL adds informativeness to K-means-based optimization for mini-batch querying.
  • Dataset and scalability considerations: Dataset characteristics and batch size determine whether diversity-based or uncertainty-based querying is more suitable.Diversity methods tend to work better with richer category content and larger batches, whereas uncertainty methods tend to work better with smaller batches and less rich content.

3.2 Data Expansion of Labeled Samples in DeepAL

DeepAL expands scarce labeled data by exploiting pseudo-labels, unlabeled data, generated samples, and alternative training regimes. These strategies can improve performance without proportionally increasing annotation costs.

  • Motivation: DeepAL addresses insufficient labeled data by enlarging training information without substantially increasing labeling costs.The survey frames data utilization and training-method exploration as essential for supporting deep model updates under limited annotation budgets.
  • Pseudo-labeling: CEAL adds high-confidence pseudo-labeled samples to the queried labeled set before training the deep model.The expanded training set combines oracle-labeled samples with model-generated labels.
  • Generative augmentation: GAAL introduces GAN-based generation into active-learning queries, while BGADL addresses the risk that random augmentation may produce uninformative samples.The passage contrasts generative augmentation with the concern that generated samples may waste computation if they do not add information.
  • Adversarial representation learning: VAAL and ARAL jointly use labeled, unlabeled, and generated data with adversarial representation learning to improve sample utilization.ARAL extends VAAL by incorporating samples produced by deep generative networks into joint model training.
  • Alternative training regimes: SSAL combines unsupervised, supervised, and semi-supervised learning across active-learning cycles, and this training method yields a surprising performance improvement.The reported improvement is described as larger than differences among sampling strategies.
  • Implications: Training-method and data-utilization changes may produce performance gains exceeding those from changing the query strategy, without additional labeling costs.These techniques are presented as a way to use existing data when queried samples are insufficient for updating deep models.

3.3 DeepAL Generic Framework

The generic DeepAL framework combines deep models for feature extraction with active learning for sample querying, while adapting uncertainty estimation to deep representations. The survey presents this framework as broadly transferable but not yet fully integrated.

  • Representative framework: CEAL combines a deep convolutional neural network with active learning by querying uncertain samples and pseudo-labeling high-confidence samples.This framework was introduced for depth image classification and separates small uncertain samples from larger high-confidence groups.
  • Generic framework: A common DeepAL pipeline trains a deep model on limited labeled data, extracts unlabeled-data features, and uses them to estimate uncertainty for querying.This framework is illustrated for applications including medical image segmentation and defect detection.
  • Uncertainty estimation: DeepAL methods use outputs from final or middle hidden layers as the basis for uncertainty or diversity estimation.The survey notes that some methods use the first fully connected layer, while others use middle hidden layers.
  • Uncertainty estimation: Traditional uncertainty-based query strategies cannot be directly applied to deep models because deep and shallow models use different learning paradigms.Figure 8 contrasts traditional uncertainty measurement with a two-stage method combining feature extraction and task learning.
  • Scope and outlook: The task-independent generic framework can be transplanted across fields, but tighter integration between deep feature extraction and active querying requires further exploration.The survey characterizes current fusion as assigning feature extraction mainly to DL and sample querying mainly to AL.

3.4 DeepAL Stopping Strategy

DeepAL stopping strategies determine when annotation querying should end. The survey describes stability-based stopping as an alternative to predefined criteria that may stop too early or too late.

  • Predefined criteria: Most DeepAL methods use predefined stopping criteria such as iteration limits, accuracy-change thresholds, or labeled-sample counts.These criteria stop querying when a preset condition is satisfied.
  • Predefined criteria: Predefined stopping criteria may prevent optimal performance because premature stopping loses performance while excessive annotation wastes budget.The survey identifies both under-querying and over-querying as practical risks.
  • Stability-based stopping: Stabilizing Predictions forms an unlabeled stop set and ends querying when model predictions on that set become stable.The stop set avoids requiring additional labels, preserving the annotation-saving objective of active learning.
  • Challenge summary: Table 1 summarizes challenges in combining deep learning and active learning together with corresponding solution approaches.The table provides a consolidated view of the framework’s challenge–solution mapping.

4 APPLICATION OF DEEPAL IN FIELDS SUCH AS VISION AND NLP

DeepAL has been applied across visual processing, language, speech, social, medical, industrial, wildlife, and disaster-related tasks. The survey organizes these applications across multiple fields and task types.

  • Vision: DeepAL applications include visual tasks such as object detection and semantic segmentation.
  • Language and analysis: DeepAL has also been applied to NLP, speech and audio processing, social network analysis, and medical image processing.
  • Specialized applications: Other reported application areas include wildlife protection, industrial robotics, and disaster analysis.

4.1 Visual Data Processing

Visual DeepAL research addresses costly labeling across image, object, segmentation, and video tasks by combining deep representations with selective sample acquisition. Applications report reduced annotation demands while maintaining or improving task performance.

  • Image classification and recognition: Image-focused DeepAL targets efficient querying of high-dimensional data at low labeling cost, where traditional active learning performs poorly.
  • Image classification and recognition: CEAL combines uncertainty-based querying with high-confidence pseudo-labels to expand the training set for image classification.
  • Medical imaging: Medical image annotation is expensive because it requires specialized expertise, motivating active learning for tissue and medical-image analysis.
  • Image classification and recognition: DeepAL applications include captcha recognition and face recognition, using deep CNN feature extraction alongside active learning to reduce labeling costs.
  • Object detection and semantic segmentation: For finger bone segmentation, an iterative deep-supervised active learning method achieved comparable segmentation results using fewer samples than complete markup.
  • Video processing: Video annotation is more expensive than image annotation because video tasks require processing temporal as well as spatial information.
  • Video processing: Video DeepAL studies report improved navigation strategies with fewer samples and activity recognition with fewer labeled samples, fewer resources, and high accuracy.

4.2 Natural Language Processing (NLP)

NLP applications use DeepAL to address expensive labeling in translation, classification, semantic analysis, information extraction, and question answering. Reported approaches reduce annotation needs or improve performance across specialized language tasks.

  • Machine translation: Low-resource machine translation requires costly parallel corpora, motivating active selection of information-source sentences for corpus construction.
  • Text classification: Text classification methods select influential embedding-space samples or use recurrent acquisition functions to reduce labeling and training costs without reducing accuracy.
  • Semantic analysis: Deep active learning supports sentiment and news accuracy detection, including a system using 1-2 orders of magnitude fewer annotation samples than fully supervised learning.
  • Semantic analysis: The news accuracy detection system reported 25% average performance gains while using substantially fewer annotation samples.
  • Information extraction: Information-extraction applications combine recurrent models, GRU-based models, transfer learning, and active learning for disaster-related tweets, entity recognition, and entity resolution.
  • Question answering: Question-answering research uses online active learning with deep models to learn incrementally from user feedback in open-domain dialogue.

4.3 Other Applications

Beyond vision and NLP, DeepAL is applied to biological data, wearable activity recognition, and speech emotion recognition. The survey characterizes these applications as broad but still early-stage, with most work concentrated in visual processing.

  • Scope of applications: DeepAL applications extend to gene expression, robotics, wearable-device analysis, social networking, and ECG signal analysis.
  • Biological applications: MLFS combines deep learning and active learning to select genes or miRNAs from expression profiles while considering biological relationships.
  • Wearable devices: Wearable and mobile-device studies use DeepAL for human activity recognition, including context-aware annotator selection and resource-efficient incremental models.
  • Speech and audio: Speech emotion recognition uses active learning to train deep neural networks when label resources are limited and generalization to new domains is difficult.
  • Scope of applications: DeepAL research remains preliminary, with applications mainly concentrated in visual image processing despite broader reported practical value.

5 DISCUSSION AND FUTURE DIRECTIONS

DeepAL combines deep learning’s feature extraction with active learning’s potential to reduce annotation costs, but its evaluation remains inconsistent and its research is still developing. Future work emphasizes unified evaluation, reproducibility, broader task independence, and mixed improvement strategies.

  • Motivation and value: DeepAL combines high-dimensional data processing and automatic feature extraction with active learning’s potential to reduce annotation costs.This potential is especially relevant where labels require high levels of expertise and are difficult to obtain.
  • Evaluation challenges: Reported DeepAL results vary substantially across studies, including a 13% difference for CIFAR 10 random sampling at 20% labeled data.The same method also differed by 8% across studies using CIFAR 100 and VGG16 at 40% labeled data.
  • Evaluation challenges: A general evaluation platform, unified high-performance random sampling baseline, shared experimental settings, repeated runs, ablations, and transfer experiments are needed.These practices are proposed to clarify inconsistent findings and identify which improvements produce performance gains.
  • Research directions: Current DeepAL research focuses on improving query strategies, optimizing training methods, and developing task-independent models.Query-strategy work increasingly considers uncertainty, diversity, and hybrid selection strategies, while training work expands labeled or unlabeled data and combines learning paradigms.
  • Research directions: Task-independent DeepAL could extend more directly across tasks, but existing work remains insufficient and often concentrates on uncertainty-based selection.The survey also cautions that pursuing smaller training subsets blindly may be unwise when sample-importance differences are negligible.
  • Future outlook: Mixed improvement strategies are identified as an important future direction, while DeepAL is characterized as practically valuable but still in its infancy.The proposed directions are described as compatible rather than mutually exclusive.

6 SUMMARY AND CONCLUSIONS

The paper presents a comprehensive and systematic review of deep active learning, addressing the challenges of combining active and deep learning. It organizes existing work by major methodological perspectives, surveys applications, and discusses future directions.

  • Scope and contribution: The survey comprehensively analyzes the necessity and challenges of combining traditional active learning with deep learning.It frames this analysis as the first comprehensive and systematic review of deep active learning.
  • Scope and contribution: Existing work is analyzed from query strategy optimization, labeled sample data expansion, and model generality perspectives.The survey also summarizes DeepAL stopping strategies.
  • Scope and contribution: The paper reviews DeepAL applications and discusses comprehensive future directions.Its application-oriented analysis complements the methodological classification of existing work.
Loading 2009.00236v2…