A Survey of Deep Active Learning

记录一篇深度主动学习的调研文章。 2020年，这篇 survey 被挂在了 arxiv 上， 见此链接。 此篇 survey 全30页，引用189篇参考文献。

Introduction

Background: A rich variety of related work has been published, DeepAL still lacks a unified classification framework.

Challenges of DeepAL:

• Insufficient data for label samples
  • The labeled training samples provided by the classic AL are insufficient to support the training of traditional DL.
  • The non-batch sample query method commonly used in AL is not applicable in the DL context.
• Model uncertainty
  • The softmax response of the final output is unreliable as a measure of confidence, and the performance of this method will thus be even worse than that of random sampling. DL model can be too confident about the output results.
• Processing pipeline inconsistency
  • AL: fixed features + learned classifiers
  • DL: learned features + learned classifiers
  • Only fine-tuning the DL models in the AL framework, or treating them as two separate problems, may thus cause divergent issues.

Approaches

In general they summarized the corresponding approaches for the challenges.

• Insufficient data for label samples
  • data augmentation
  • assigning pseudo-labels
  • combine supervised and semi-supervised training
  • batch sample
• Model uncertainty
  • Bayesian deep learning
• Processing pipeline inconsistency
  • Modify the combined framework of AL and DL to make the proposed DeepAL model as general as possible

Query strategies

Batch Mode DAL should be satisfied. The reference indexes are copied.

1. Uncertainty-based and hybrid query strategies
   • directly use uncertainty sampling (non-batch) [9, 59, 114, 123]
   • query batch sample set representative to the distribution [177:Exploration-P, 183:DBAL, 99:WI-DL]
     • the richer the category content of the dataset, the larger the batch size, and the better the effect of diversity-based methods;
   • query batch sample set representative to the gradient embedded space [10:BADGE]
   • query batch sample set representative by adversarial method [146:WWAL, 150:VAAL, 81:TA-VAAL]
2. Deep Bayesian Active Learning
   • obtain the posterior distribution of network prediction [47:DBAL, 118:DEBAL, 28:DPEs, 114:ActiveLink]
   • obtain the mutual information between the batch samples and the model parameters [84: BatchBALD]
   • reconstructed the batch structure to optimize the sparse subset approximation [117:ACS-FW]
   • other [54, 105, 126, 147, 175, 179]
3. Density-based Methods
   • core-set approach [49:FF-Active, 138:core-set]
   • discriminative approach [35:Active Palmprint Recognition, 51:DAL]
4. Other
   • RL approach [37]
   • Adversarial examples[36:DFAL]
   • ensemble [14]
   • flexible acquisition function [57:RAL, 100:DRAL]
   • with NAS [50:Active-iNAS]

Model training

Normally accompanied with insufficient data (under the view of DAL).

1. assigning pseudo-labels [166:CEAL]
2. combine unsupervised feature learning [99:WI-DL]
3. data augmentation [187:GAAL, 162:BGADL, 107:ARAL]
4. add adversarial task [150:VAAL, 81:TA-VAAL]

DAL on the generalization of the model

1. Use the final layer output to make query
2. ALso use the middle layer output to make query [59:AL-MV, 178:LLAL, 182]