A Survey of Deep Active Learning (2009.00236v2)

Abstract: 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.

A Survey of Deep Active Learning

The paper "A Survey of Deep Active Learning" by Pengzhen Ren et al. provides a comprehensive examination of the intersection between Deep Learning (DL) and Active Learning (AL), culminating in a field known as Deep Active Learning (DeepAL). This survey endeavors to bridge the gap in the literature regarding a systematic classification and review of existing works within this emergent domain.

Overview of DeepAL

Motivation and Challenges

DeepAL, emerging from the synergy of DL and AL, presents a compelling approach to leveraging DL’s powerful feature extraction capabilities while potentially reducing the annotation costs associated with AL. The principal motivations for DeepAL are:

1. High Dimensional Data Handling: DL's efficacy in managing high-dimensional data is unparalleled, a feature that traditional AL methods lack.
2. Cost Reduction in Annotation: AL aims to minimize labeled sample requirements, positing DeepAL as a cost-effective solution for domains where data annotation is labor-intensive and expensive, such as medical imaging and speech recognition.

However, integrating DL with AL into a cohesive framework poses several challenges:

• Model Uncertainty: Traditional uncertainty-based query strategies in AL do not directly translate to DL, necessitating nuanced approaches to accurately gauge model certainty.
• Sample Scarcity: DL's data greediness contrasts AL’s paradigm of limited sample labeling, leading to training inefficiencies.
• Pipeline Discrepancies: Combining the pipelines of AL and DL requires a paradigm shift to avoid divergence issues between feature learning and sample querying.

Framework and Methodologies

Batch Mode DeepAL (BMDAL)

BMDAL underpins DeepAL by replacing the inefficient one-by-one query strategy of traditional AL with batch-based sample querying. The core idea is to maximize the mutual information between batch samples and model parameters to ensure both informational richness and diversity within queried batches. Methods like BatchBALD utilize joint mutual information mechanisms to achieve this goal.

Query Strategy Optimization

The survey identifies several strategies within DeepAL:

1. Uncertainty-based and Hybrid Query Strategies: These strategies rank samples based on uncertainty or combine multiple attributes (e.g., informativeness and diversity) to select high-value samples effectively. For instance, exploration/exploitation-based strategies balance these aspects to optimize sample utility.
2. Deep Bayesian Active Learning (DBAL): DBAL harnesses Bayesian methods, often using techniques like Monte-Carlo dropout, to estimate model uncertainty and apply it within AL frameworks effectively.
3. Density-based Methods: These approaches, such as core-set selection, focus on representative subsets that preserve the overall data distribution, addressing annotation cost constraints.

Data Expansion for Labeled Samples

To mitigate the labeled sample scarcity, the paper reviews methods such as pseudo-label assignment for high-confidence predictions (e.g., CEAL) and the incorporation of generative models (e.g., GANs) to augment training datasets, thereby enhancing model learning from a broader data spectrum without incurring additional annotation costs.

Generic Frameworks and Task Independence

The synthesis of AL and DL into a coherent framework involves:

• Adaptive Query Strategy Integration: Approaches like LLAL leverage multi-view uncertainty measurements across DL model layers, promising more refined selection criteria.
• Incremental Training: Continual learning paradigms alleviate the computational load by incrementing training datasets rather than retraining models from scratch in each AL cycle.

Application Domains

DeepAL exhibits extensive applicability across diverse domains:

• Visual Data Processing: Encompasses image classification, object detection, and video processing. Significant strides are noted in biomedical imaging, autonomous navigation, and industrial defect detection.
• NLP: Tasks include machine translation, text classification, and semantic analysis, where DeepAL methods significantly curtail annotation costs while preserving model efficacy.
• Other Fields: Applications span from gene expression analysis and robotics to social network analysis and wearable device data analytics.

Future Directions

The survey elucidates potential research trajectories for DeepAL:

• Unified Evaluation Platforms: Establishing standardized benchmarks to evaluate and compare DeepAL methodologies is crucial.
• Enhanced Hybrid Strategies: Fusing diverse and uncertainty-based strategies can yield robust sample selection mechanisms.
• Incremental and Semi-Supervised Learning: Developing methods to leverage unlabeled data effectively and increment model training represent pivotal growth areas.

Conclusion

Pengzhen Ren et al.’s survey of Deep Active Learning delivers a nuanced and detailed overview of integrating Deep Learning with Active Learning. By systematically categorizing existing methodologies, delineating application scenarios, and proposing future research avenues, this paper lays a solid foundation for advancing the field of DeepAL, driving its adoption in practical, high-stakes domains where annotation costs are a pervasive constraint.