Overview of Active Learning (ActiveLearning)

I. Introduction

This article summarizes my understanding of active learning and my insights into the latest research. The main purpose is for everyone to refer to, discuss, and learn and exchange active learning techniques together. At the same time, I will continue to read articles in the field of active learning in the future. If there are good and recommended articles, I will update them in my github in real time. You can read the latest and most important articles as quickly as possible through this list. You are also welcome. Recommend some articles to me and communicate with me.

Awesome Active Learning:

https://github.com/baifanxxx/awesome-active-learning

Note: The first 1, 2, and 3 sections are all about basic active learning content. There are many articles that have similarly organized and introduced them. If you already know a lot about it, you can jump directly to section 4 and beyond.

2. Introduction

Active learning is a machine learning or artificial intelligence method that actively selects the most valuable samples for labeling. The purpose is to use as few, high-quality sample annotations as possible to achieve the best possible performance of the model. In other words, the active learning method can improve the gain of samples and labeling, and maximize the performance of the model under the premise of limited labeling budget. It is a solution to improve data efficiency from the perspective of samples, so it is applied in situations where labeling costs are high. , tasks with difficult annotation, such as medical images, driverless driving, anomaly detection, and related issues based on Internet big data.

▲ The pool-based active learning cycle. (Burr Settles, 2010)

Settles, Burr's Active Learning Literature Survey article provides a summary of classic active learning work. The picture above is a classic pool-based active learning framework. In each active learning cycle, based on the information of the task model and unlabeled data, the query strategy selects the most valuable samples and gives them to experts for labeling and adds them to the labeled data set to continue training the task model. Because there are human annotations in the process of active learning, active learning is also a type of Human-in-the-Loop Machine Learning.

Why is active learning useful? Let everyone feel it through a small intuitive example.

▲ (a) A data set consisting of 400 instances uniformly adopted from two classes of Gaussian distributions. Instances are represented as points in a two-dimensional feature space. (b) Randomly sample 30 labeled instances from the problem domain and train a logistic regression model. This blue line represents the classifier's decision boundary (70% accuracy). (c) Logistic regression model (90%) trained on 30 active query instances using uncertainty sampling. (Burr Settles, 2010)

This shows that the contribution of samples to the model is not the same, and it is of practical significance to select more valuable samples. Of course, how to determine and evaluate the value of samples is also a focus of active learning research.

3. Model classification

According to application scenarios, active learning methods can be divided into three types: membership query synthesis, stream-based and pool-based. Among them, pool-based is the most common scenario, and because deep learning is based on the batch training mechanism, the pool-based method is easier to fit with it.

In the scenario of membership query synthesis, the algorithm may select any part of the entire unlabeled data and hand it to oracle for labeling. The typical assumption is that it includes the data generated by the algorithm itself. But sometimes, the data generated by the algorithm cannot be recognized by Oracle. For example, the generated handwritten image is too strange, and Oracle cannot recognize that it belongs to 0~9? Or the generated audio data does not have semantic information, making it impossible for Oracle to recognize it.

In the stream-based scenario, only one unlabeled sample is input to the algorithm each time, and the algorithm decides whether to hand it over to the oracle for labeling or reject it directly. It's a bit like a defective product inspector on an assembly line. When a product comes over, it needs to be judged immediately whether it is a defective product, and it cannot be considered based on the overall situation of this batch of products at the beginning.

In the pool-based scenario, a batch of unlabeled samples is input to the algorithm each time, and then the algorithm selects one or several samples based on the strategy and delivers them to the oracle for labeling. Such scenes are more likely to occur in life, and algorithms can also be compared and comprehensively considered based on this batch of samples.

4. Basic query strategy

In the active learning framework, the most important thing is how to design a query strategy to judge the value of the sample, that is, whether it is worthy of being labeled by the oracle. The value of a sample is not static. It is not only related to the sample itself, but also to factors such as tasks and models. A simple example, in the classification problem of cats and dogs, a photo of a dog that looks like a cat is often valuable for training the classification model because it is difficult to distinguish.

However, when the same photo appears in the animal and plant classification problem, it becomes less important, because it is not difficult for the model to distinguish it. Therefore, the design of query strategies is not simple and static, and needs to be set according to specific circumstances, problems, and needs. This has resulted in a variety of query strategies. Below, I introduce some basic query strategies for your reference.

Uncertainty Sampling:Perhaps the simplest, most straightforward and most commonly used strategy. The algorithm only needs to query the most uncertain samples to label the oracle. Under normal circumstances, the model can quickly improve its performance by learning the labels of samples with high uncertainty. For example, when students are answering questions, they will definitely improve faster if they only do the questions they are prone to make mistakes than if they randomly select some questions. For some models that can predict probabilities, such as neural networks, probabilities can be directly used to represent uncertainty. For example, directly use the probability value, the difference between the first and second probability values, the entropy value, etc.

Diversity Sampling: is a common strategy that considers the distribution of data. The algorithm ensures that the queried samples can cover the entire data distribution according to the data distribution to ensure the diversity of the annotated data. For example, when setting exam questions, teachers will try their best to come up with some representative questions, and at the same time try to ensure that every chapter is covered, so as to ensure the diversity of questions and comprehensively examine the students' comprehensive level. Similarly, among the methods adopted by diversity, they are mainly divided into the following methods:

Model-based outliers—use outlier samples that give the model low activation because the existing data lacks this information;

Representative Sampling - Select some of the most representative samples, such as using clustering and other methods to obtain representative samples and find representative samples based on differences in different domains ;

Real Scene Diversity—Sampling fairly based on the diversity and sample distribution of real scenes.

Expected Model Change:EMC usually selects the sample that has the greatest change and greatest impact on the current model to label oracle. Generally speaking, it needs to be based on the label of the sample. Back propagation calculates the changes or gradients of the model, etc. In practical applications, in order to weaken the premise of requiring labels, the prediction results of the model are generally used as pseudo labels and then the expected model changes are calculated. Of course, there are certain problems with this approach. The pseudo labels and the real labels are not always consistent, which is related to the prediction performance of the model.

Query-By-Committee:QBC uses a committee composed of multiple models to vote on the candidate data, that is, make decisions separately, and finally they choose the most suitable The divergent samples are labeled as the most informative data to the oracle.

In addition, some researchers combine multiple query strategies and use hybrid strategies to query, for example, taking into account both uncertainty and diversity. There are some other query strategies, such as expected error reduction, variance reduction, density weighting, etc.

5. Classic method

Below I will share with you some classic active learning methods, which are often compared. In your future articles, you can also consider comparing with the following classic methods.

Entropy

The entropy value can be calculated directly based on the predicted probability distribution, and the sample with the largest entropy value can be selected for labeling.

BALD

Deep Bayesian Active Learning with Image Data

https://arxiv.org/abs/1703.02910

BGADL

Bayesian Generative Active Deep Learning

https://arxiv.org/abs/1904.11643

Core-set

Active Learning for Convolutional Neural Networks: A Core-Set Approach

https://openreview.net/forum?id=H1aIuk-RW

LLAL

Learning Loss for Active Learning

https://arxiv.org/abs/1905.03677?context=cs.CV

VAAL

Variational Adversarial Active Learning

https://arxiv.org/abs/1904.00370

6. Application scenarios

Since active learning solves the problem of how to select high-value samples from unlabeled data for labeling, it is widely used in scenarios and practical problems where data labels are difficult to obtain and labeling costs are high.

Internet big data related applications:In the Internet big data scenario, there are countless unlabeled data, but it is impossible to label all the data. With limited funds and time, the most effective method is to use active learning to select the most valuable samples and hand them over to humans for labeling. For example,

Alibaba shopping technology

https://www.zhihu.com/question/265479171/answer/1495497483

Zhongke Zhiyun launches a new active learning algorithm framework in the world, subverting the traditional large-scale sample and manual labeling model

https://www.ofweek.com/ai/2021-07/ART-201713-8210-30509389.html

In fields such as security risk control anomaly detection, abnormal data is far less than normal data, and it is extremely unreasonable to label a large amount of data on the network. However, active learning can selectively label these data.

Applications in robotics fields such as driverless driving:In many robotics fields, it is necessary to collect a large amount of labeled data for training. Especially the very hot field of autonomous driving. In the field of autonomous driving, the perception of the environment by autonomous vehicles is particularly important. The quality of perception directly affects the quality of decision-making and plays a vital role in the safety of autonomous vehicles.

Perceptual models are mostly built using deep learning, and the importance of data is self-evident, especially annotated data. There are many types and complex unmanned driving environments in real scenes. In order to ensure the performance of the model, most companies need to use images, point clouds and other data collected by running the car in the actual scene for training. Faced with such a huge amount of data, it is almost an impossible task to label every sample. Instead, use active learning to select the most valuable samples (which may be due to the high uncertainty of the current model predictions) and then manually label them. Continue Train the model to maximize model performance, improve stability and security. For example, Tesla etc.

Tesla challenges visual limits

https://www.bilibili.com/read/cv7621643

How Active Learning Improves Nighttime Pedestrian Detection for Autonomous Driving [NVIDIA]

https://www.bilibili.com/video/BV1xV411o72V/

What’s so special about the training systems behind Waymo and Tesla?

https://zhuanlan.zhihu.com/p/400834629

Intelligent medical diagnosis and other fields:In the medical field, the development of deep learning has brought revolutionary developments in many aspects, including diagnosis. Data-driven methods inevitably require a large amount of labeled data, and labeling medical images is not only time-consuming and labor-intensive, but also requires specific professional knowledge. Therefore, it is of great practical significance to use active learning to select samples that are difficult to predict by the model for selective labeling.

There are many papers studying the application of active learning in the medical field, but in actual application and implementation, the most important issues facing medical diagnosis are accuracy and generalization performance. Since medical data consists of small samples, these most important and basic issues have not been completely solved, so active learning is not very popular. However, there are still some companies applying it. For example, Tencent AI Lab uses active learning and difficult example mining solutions.

China's first smart microscope has been approved to enter clinical practice: pathological diagnosis is AI-based, built by Tencent AI Lab

https://new.qq.com/omn/20200409/20200409A0BGWI00.html

In my opinion, if the amount of medical data obtained is small, there is no need to apply active learning, because with limited samples, it is difficult to achieve satisfactory performance even if they are all labeled, let alone make a selection. But the real demand scenario is,

1. There is a large amount of unlabeled medical data, from which valuable ones need to be selected for labeling, such as labeling images from video data (gastrointestinal endoscope videos) for detection, etc.;

2. To truly achieve basic performance and deploy it, you still need to collect data and label it during use for a long time. However, since this process is continuous and long-term work, it is also necessary to proactively deal with such a large amount of unlabeled data. Learn to select annotations.

In short, the application scenario of active learning is about how to save the labeling workload and achieve satisfactory performance of the model when there is a large amount of unlabeled data (at least not lacking). In the era of deep learning explosion, various tasks and applications are considering using data-driven learning methods to solve them, which requires higher data. In practical applications, it is impossible to completely abandon labels or unlabeled data. Active learning can provide a more reasonable expedient. It is necessary to label valuable data without labeling all the data selectively. Label.

7. Possible problems in practical application

Although it is relatively practical considering the starting point of active learning and the problems to be solved, there are still some problems in the practical application of current active learning methods.

Unstable performance:The biggest problem restricting active learning is unstable performance. Active learning selects from samples according to a self-specified selection strategy. In this process, the strategy and data samples are two very important factors that affect performance. For very redundant data sets, active learning is often better than random sampling. However, for data sets with very diverse sample data and low redundancy, active learning sometimes has worse effects than random sampling. The distribution of data samples also affects different active learning methods, such as uncertainty-based methods and diversity-based methods, which have inconsistent effects on different data sets. This instability of performance is one of the constraints that restricts people from applying active learning. Key factor.

In practical applications, data selection and labeling need to be performed based on active learning. If the strategy at this time is worse than random sampling, people cannot change or stop losses in time because the data has already been labeled and sunk costs have already been incurred. However, these methods of optimizing network structure and performance do not have this problem. People can always try different methods and techniques to achieve the best performance, and the loss of modifications and attempts is very small.

Active learning is even more demanding. It almost requires work to directly apply the designed strategy. If you don't work, the selected samples will still be labeled, and you will still lose time and money. The stringent requirements and unstable performance lead people to save this energy and directly use random annotation.

Challenges of dirty data:Now almost all papers are tested and studied on public data sets and ready-made data sets. In fact, these data sets have been selected and filtered, extreme outliers have been removed, and sample balance has even been taken into consideration, artificially labeling categories with few samples more and categories with many samples less. In actual applications, the data situation is far from this ideal data set. Active learning is a commonly used uncertainty selection strategy. It is not difficult to imagine that noisy samples or even outliers will always be selected and labeled. Such samples may not only not improve the performance of the model, but may even worsen the performance.

In practice, there is also the problem of OOD (out of distribution). For example, if you want to train a cat and dog classifier, you can directly search for cats and dogs from the Internet by keyword and collect a large number of pictures. There may be some tigers, lions, wolves, etc. that are not in the cat and dog category. irrelevant samples, but their uncertainty is very high, and if selected, it will not improve the performance of the model.

Difficult to transfer:Active learning is a data selection strategy, so practical applications inevitably require a more versatile and better generalizable active learning strategy. However, the current active learning strategy is difficult to transfer between different domains and different tasks. For example, an active learning strategy for cat and dog classification tasks was designed, which achieved good performance based on uncertainty or diversity. Now we need to make a For the new chicken and duck classification task, is it necessary to redesign a strategy? What if the task is the classification of diseased tissue?

Since the data distribution characteristics of different tasks may be different and the difficulty of different tasks is different, there is no guarantee that the active learning strategy can be universally used in different tasks with different data. It is often necessary to design an active learning strategy for fixed tasks. This consumes energy. If there is a versatile active learning strategy, it can be transferred to different tasks and applied more widely. It can even be directly deployed as a general annotation software for various tasks and data sets. Provides active selection and labeling functions.

Inconvenient interaction:The data selection strategy is closely related to the annotation process. The ideal process is to have an integrated software that can provide active data selection and then provide an interactive interface for annotation. This It is to combine the active learning process with annotation software. Having only efficient active learning strategies without convenient annotation interaction will also cause additional waste of energy. In terms of process, active learning now generally selects a batch of samples to be labeled and gives them to people to label. It is expected that people can label them as soon as possible and hand them over to the model, and the model will continue to train and then select again.

When people annotate, the model cannot be trained, actively learns, or performs other operations. It is a serial process. You need to wait for the manual annotation to complete before proceeding with the next training. Such a process is not so convenient and efficient. Imagine applying the active learning + labeling system to doctors. The strategy first selects some samples, and the doctor only labels these samples for a few days, and then trains the model. The model is trained for a period of time. After a while, some samples are selected and given to the doctor. The doctor and the model wait for each other's operations, which reduces efficiency and convenience.

8. Latest research directions and paper recommendations

Below I will introduce some of the latest papers with high reading value for active learning, so that everyone can grasp the research directions and hot spots. If you are interested, you can continue to follow my awesome-active-learning paper list on github. I will update valuable active learning work in real time for everyone to learn and communicate.

8.1 Exploration of active learning issues and methods

There are still some shortcomings in the current basic methods and problems of active learning. There are some latest methods trying to solve these problems.

Mind Your Outliers! Investigating the Negative Impact of Outliers on Active Learning for Visual Question Answering (Co-author, Li Feifei)

https://arxiv.org/abs/2107.02331

Active learning promises to alleviate the massive data requirements of supervised machine learning: it has successfully improved sample efficiency by an order of magnitude for traditional tasks such as topic classification and object recognition. However, the authors found that in stark contrast to this phenomenon, a variety of active learning methods failed to outperform random selection across 5 models and 4 datasets on the visual question answering task. To understand this difference, the authors analyzed 8 active learning methods on a per-example basis and identified the problem as collective outliers—a set of examples that the active learning method prefers to acquire but that the model cannot learn (e.g., ask Problems with the text are in the image or require external knowledge).

Through systematic ablation experiments and qualitative visualizations, the authors verify that collective outliers are a common phenomenon that leads to the degradation of pool-based active learning. Notably, the authors show that active learning sample efficiency increases significantly as the number of collective outliers in the active learning pool decreases.

Contrastive Coding for Active Learning Under Class Distribution Mismatch：

https://openaccess.thecvf.com/content/ICCV2021/html/Du_Contrastive_Coding_for_Active_Lea

Active learning (AL) is successful based on the assumption that labeled and unlabeled data are obtained from the same distribution. However, its performance deteriorates in the case of class distribution mismatch, where the unlabeled data contains many samples outside the class distribution of the labeled data. In order to effectively deal with the AL problem under class distribution mismatch, the author proposes an AL framework based on contrastive coding, named CCAL.

Unlike existing AL methods that focus on selecting the most informative samples for labeling, CCAL extracts semantic and unique features through contrastive learning and combines them in a query strategy to select the most informative unlabeled samples with matching categories. sample. Theoretically, the authors demonstrate that the AL error of CCAL has a strict upper bound.

LADA: Look-Ahead Data Acquisition via Augmentation for Active Learning：

https://arxiv.org/abs/2011.04194

The potential benefits of virtual instances generated from data augmentation have not been considered in the acquisition process of active learning. During the data acquisition process, data augmentation selects and generates data instances that are informative for training the model. Therefore, the authors propose the integration of data acquisition and data augmentation through augmentation or prospective data acquisition with LADA. Before the acquisition process, LADA considers 1) selected unlabeled data instances and 2) virtual data instances generated through data augmentation. Furthermore, to enhance the informativeness of virtual data instances, LADA optimizes data augmentation strategies to maximize prediction acquisition scores, resulting in the proposals of InfoMixup and InfoSTN. Since LADA is a generalizable framework, the authors experimented with various combinations of acquisition and enhancement methods.

8.2 Combination of active learning and semi-supervised learning

Since semi-supervised learning has shown excellent performance, if active learning and semi-supervised learning can be combined when there are insufficient labels, even better performance will be achieved.

Semi-Supervised Active Learning for Semi-Supervised Models: Exploit Adversarial Examples With Graph-Based Virtual Labels：

https://openaccess.thecvf.com/content/ICCV2021/html/Guo_Semi-Supervised_Active_Learnin

Although current mainstream methods are beginning to combine SSL and AL (SSL-AL) to mine diverse representations of unlabeled samples, fully supervised task models of these methods are still trained using labeled data only. Furthermore, there is a mismatch problem in the SSL-AL framework of these methods. Here, the authors propose a graph-based SSL-AL framework to unleash the power of SSL models and conduct efficient SSL-AL interactions.

In this framework, SSL utilizes graph-based label propagation to provide pseudo-labels for unlabeled samples, render the structural distribution of AL samples and boost AL. AL finds samples near the decision boundary, leveraging adversarial examples to help SSL perform better label propagation. The information exchange in the closed loop realizes the mutual enhancement of SSL and AL.

8.3 Combination of active learning and unsupervised domain adaptation

Unsupervised domain adaptation requires aligning the target domain and the source domain, so that the model can use the data and labels of the source domain to achieve better performance in the unlabeled target domain. There are currently some works that consider the relationship between the source domain and the target domain and design active learning strategies to improve the performance of the model in the target domain.

Multi-Anchor Active Domain Adaptation for Semantic Segmentation：

https://arxiv.org/abs/2108.08012

Unconditionally aligning the target domain distribution with the source domain may distort information unique to the target domain data. To this end, the authors propose a novel multi-anchor-based active learning strategy to assist in domain-adaptive semantic segmentation tasks. By innovatively employing multiple points instead of a single centroid, the source domain can be better characterized as a multimodal distribution, and the practice selects more representative and complementary samples from the target domain. Manually annotating these samples requires very little effort and can effectively alleviate the distortion of the target domain distribution, thereby achieving large performance gains. A multi-anchor strategy is also used to model the target distribution. Regularizing the latent representation of compact target samples around multiple anchor points through a soft alignment loss enables more accurate segmentation.

8.4 Combination of active learning and knowledge distillation

In the process of knowledge distillation, teachers transfer knowledge to students, but what kind of samples can help this process is also a direction that active learning can research.

Active Learning for Lane Detection: A Knowledge Distillation Approach：

https://openaccess.thecvf.com/content/ICCV2021/html/Peng_Active_Learning_for_Lane_Detection_A_

The authors found that existing active learning methods perform poorly in lane detection for two reasons. On the one hand, most methods evaluate data uncertainty based on entropy, which is undesirable in lane detection since it encourages the selection of images with few or even no lanes. On the other hand, existing methods are not aware of lane labeling noise, which is caused by severe occlusions and unclear lane markings.

In this paper, the authors build a novel knowledge distillation framework and evaluate image uncertainty based on the knowledge learned by the student model. The authors show that the proposed uncertainty measure overcomes the above two problems. In order to reduce data redundancy, the authors studied the influence set of image samples and proposed a new diversity measure. Finally, the authors propose a greedy algorithm for data selection by combining uncertainty and diversity metrics.

8.5 Combination of active learning and comparative learning

Comparative learning has gained momentum recently. Recently, active learning and contrastive learning have been combined to solve the problem of contrastive learning. You can appreciate it.

Active Contrastive Learning of Audio-Visual Video Representations：

https://arxiv.org/abs/2009.09805

Contrastive learning has been shown to generate generalizable representations of audio and visual data by maximizing the lower bound of the mutual information (MI) between different views of an instance. However, obtaining a strict lower bound requires an exponential sample size in MI and therefore a large number of negative samples. We can incorporate more samples by building a large queue-based dictionary, but even with a large number of negative samples, there is a theoretical limit to the performance improvement.

The authors hypothesize that random negative sampling leads to highly redundant dictionaries, resulting in suboptimal representations for downstream tasks. In this paper, the authors propose an active contrastive learning method that builds an actively sampled dictionary containing diverse and information-rich samples, thereby improving the quality of negative samples and improving the amount of mutual information in the data. performance on tasks such as video classification.

8.6 Active learning using reinforcement learning

Reinforced active learning for image segmentation：

https://arxiv.org/abs/2002.06583

Learning-based semantic segmentation methods have two inherent challenges. First, obtaining pixel-level labels is expensive and time-consuming. Second, realistic segmentation datasets are highly imbalanced: some classes are much richer than others, biasing performance toward the most representative classes. In this paper, the authors are interested in a pooling-based approach to manual labeling effort that minimizes this effort while maximizing the performance of the segmentation model on the test set. The authors propose a new active learning strategy for semantic segmentation based on deep reinforcement learning (RL).

The agent learns a strategy to select a small set of informative image regions from a pool of unlabeled data for labeling. Region selection decisions are made based on the predictions and uncertainties of the trained segmentation model. The authors' approach proposes a DQN for active learning, making it adaptable to large-scale semantic segmentation problems. The author tested on CamVid and the large-scale dataset Cityscapes.

On Cityscapes, the authors' deep reinforcement learning-based region-based DQN method requires approximately 30% less additional labeled data than the most competitive baseline at the same performance. Furthermore, we find that our method selects more underrepresented class labels compared to baselines, thereby improving their performance and helping to mitigate class imbalance.

8.7 Active learning in point clouds

Point cloud annotation is longer and more labor-intensive than image annotation, especially pixel-level point cloud annotation. Recently, active learning has gradually emerged in point cloud work, and the results are amazing and worth looking forward to. Below I introduce a representative work on point cloud semantic segmentation.

ViewAL: Active Learning with Viewpoint Entropy for Semantic Segmentation：

https://arxiv.org/abs/1911.11789

The authors propose ViewAL, a novel active learning strategy for semantic segmentation that exploits viewpoint consistency in multi-view datasets. The authors' core idea is that inconsistency in model predictions across viewpoints provides a very reliable measure of uncertainty and encourages models to perform well regardless of the viewpoint of the observed object.

To incorporate this uncertainty measure, the authors introduce a new viewpoint entropy formulation, which is the basis of our active learning strategy. In addition, the authors propose superpixel-level uncertainty calculation, which exploits the local information inherent in the segmentation task and directly reduces the annotation cost. The joint use of viewpoint entropy and superpixels effectively selects highly informative samples.

8.8 Active learning in target detection

Recently, some active learning articles have begun to be integrated into target detection and define strategies for target detection. Target detection not only involves classification but also positioning, so the definition and modeling of uncertainty in images are more diverse, making it easier to innovate your own active learning methods. You can try to do this work.

Multiple instance active learning for object detection：

https://arxiv.org/abs/2104.02324

Despite substantial progress in active learning for image recognition, there is still a lack of instance-level active learning methods specified for object detection. In this paper, the authors propose multi-instance active object detection (MI-AOD), which selects the most informative images for detector training by observing instance-level uncertainty. MI-AOD defines an instance uncertainty learning module that exploits the difference between two adversarial instance classifiers trained on the labeled set to predict instance uncertainty for the unlabeled set.

MI-AOD treats unlabeled images as bags of instances, considers feature anchors in images as instances, and estimates the uncertainty of the image by reweighting the instances in a multi-instance learning (MIL) manner. Iterative instance uncertainty learning and reweighting help suppress noisy instances, bridging the gap between instance uncertainty and image-level uncertainty.

9. Summary

All in all, active learning still has many points that can be further studied, including but not limited to:

1. Start from the basic theories and issues of active learning, improve and improve;

2. Combine with other learning methods or concepts to improve active learning or this method, such as semi-supervised, domain adaptation, knowledge distillation, reinforcement learning, etc.;

3. Apply to new backgrounds and tasks (there are few papers combined with active learning), such as point cloud classification and segmentation, medical images, target detection, etc.

4. ......

It can not only improve on the existing methods, but also design its own active learning strategies for new specific tasks and specific problems. No matter which aspect it is, active learning has many points of continued research.

Active learning has both important application value and some problems. It is a point that can be studied by both academia and industry. I hope that friends who are interested in active learning can have more discussions and exchanges together. In the future, I will continue to update you on some excellent work in the field of active learning on Zhihu and awesome-active-learning.