Update to 2022-12-17.

contents/AL_combinations.md

@@ -64,9 +64,13 @@ Reducing the labeling cost is a common need in many research fields.
 CV is quite a wide conception.
 Here we only post several subtypes in the fields.
 
+**Survey** for the whole field:
+- Deep Active Learning for Computer Vision: Past and Future [2022]
+
 Image classification:
 - [Deep active learning for image classification [ICIP, 2017]](https://ieeexplore.ieee.org/abstract/document/8297020).
 - [The power of ensembles for active learning in image classification [2018, CVPR]](https://openaccess.thecvf.com/content_cvpr_2018/papers/Beluch_The_Power_of_CVPR_2018_paper.pdf)
+- MoBYv2AL: Self-supervised Active Learning for Image Classification [2022, BMVC]
 
 Image Semantic Segmentation：
 - Geometry in active learning for binary and multi-class [2019, Computer vision and image understanding]
@@ -118,6 +122,7 @@ Object Detection:
 - Entropy-based Active Learning for Object Detection with Progressive Diversity Constraint [2022, CVPR]
 - Weakly Supervised Object Detection Based on Active Learning [2022, NPL]
 - Active Learning Strategies for Weakly-Supervised Object Detection [2022]
+- MUS-CDB: Mixed Uncertainty Sampling with Class Distribution Balancing for Active Annotation in Aerial Object Detection [2022]
 
 Point Cloud Semantic Segmentation：
 - Label-Efficient Point Cloud Semantic Segmentation: An Active Learning Approach [2021, CVPR]
@@ -485,6 +490,7 @@ For example, the ImageNet was crawled from image databases without considering s
 
 - Federated Active Learning (F-AL): an Efficient Annotation Strategy for Federated Learning [2022]
 - LG-FAL : Federated Active Learning Strategy using Local and Global Models [2022, ICML workshop]
+- Knowledge-Aware Federated Active Learning with Non-IID Data [2022]
 
 ## Hedge
 

contents/MTAL.md

@@ -30,7 +30,7 @@ In this chapter, we would only list the works which are not included in MLAL.
   (10)
 - Active Learning from Peers [2017, NeurIPS]:
   A stream-based scenario.
-- Safe Active Learning for Multi-Output Gaussian Processes [2022]: 
+- Safe Active Learning for Multi-Output Gaussian Processes [2022, ICAIS]: 
   Multi-output regression.
 
 ### Heterogeneous MTAL
@@ -45,7 +45,7 @@ In this chapter, we would only list the works which are not included in MLAL.
   They focused on the Genre Classification and Collaborative Filtering and proposed G+CTR. 
   Two tasks shares a same parameter.
   The AL strategy selects the instance which would lead to the maximum expected change of the shared parameter.
-- Multi-task Active Learning for Pre-trained Transformer-based Models [2022]:
+- Multi-task Active Learning for Pre-trained Transformer-based Models [2022, TACL]:
   Explore various multi-task selection criteria in three realistic multi-task scenarios.
 - PartAL: Efficient Partial Active Learning in Multi-Task Visual Settings [2022]:
   Only provide annotations for several tasks of the selected instances.

contents/practical_considerations.md

@@ -17,7 +17,7 @@ So if you have any comments and recommendations, pls let me know.)*
   - [Biased data](#biased-data)
   - [Cost-sensitive case](#cost-sensitive-case)
   - [Noised data](#noised-data)
-  - [Class mismatch & Out-of-distribution data](#class-mismatch--out-of-distribution-data)
+  - [Class mismatch \& Out-of-distribution data](#class-mismatch--out-of-distribution-data)
   - [Subjective labels](#subjective-labels)
   - [Logged data](#logged-data)
   - [Feature missing data](#feature-missing-data)
@@ -52,6 +52,7 @@ So if you have any comments and recommendations, pls let me know.)*
 - [The consideration of the reliability](#the-consideration-of-the-reliability)
   - [Reusablility](#reusablility)
   - [Robustness](#robustness)
+- [The consideration of the privacy](#the-consideration-of-the-privacy)
 - [The considerations of more assumptions](#the-considerations-of-more-assumptions)
   - [Include model selection](#include-model-selection)
   - [Select for evaluation](#select-for-evaluation)
@@ -429,6 +430,12 @@ The trained model is expected to be robust to the adversarial examples.
 - Towards Exploring the Limitations of Active Learning: An Empirical Study [2021]:
   They discovered the robustness to adversarial attacks and model compressions.
 
+# The consideration of the privacy
+
+Avoid to upload tons of unlabeled data to the cloud for selection.
+- Responsible Active Learning via Human-in-the-loop Peer Study [2022]:
+  Build local peer student to select instances, then pass them to the cloud.
+
 # The considerations of more assumptions
 
 ## Include model selection

on_build/AL_analysis.md

@@ -6,4 +6,13 @@
   The common explanation is that AL methods favor hard-to-learn samples, often simply called outliers, which therefore neglects the potential benefits from AL.
 
 ### Margin
-- Is margin all you need? An extensive empirical study of active learning on tabular data [2022]
+- Is margin all you need? An extensive empirical study of active learning on tabular data [2022]
+
+## Poor performance
+
+- Uniform versus uncertainty sampling: When being active is less efficient than staying passive [2022]: 
+  Prove for logistic regression that passive learning outperforms uncertainty sampling even for noiseless data and when using the uncertainty of the Bayes optimal classifier.
+
+## Fair comparison in practice
+
+- Randomness is the Root of All Evil: More Reliable Evaluation of Deep Active Learning [2022]