Most supervised machine learning models require to be trained with large amounts of data to get good results. In most cases, these large datasets are unlabeled since the data is too large to be labeled manually. It requires enormous human effort and time, and we can end up with problems like missing or erroneous annotations, which can affect the training of the model and, hence, further impact the performance of the model. But here's the good news: active learning is here to save the day! Basically, it's a way to cut down on how much we have to label by making use of all the other examples we have lying around that we haven't labeled.

Active Learning: Getting Machines Smarter With Less Work

Active learning is used to construct a high-performance algorithm by actively selecting relevant data points for labeling. While training, active learning prioritizes the data which needs to be labeled in order to train a supervised learning model better. It does that by smartly identifying the labels which would be the most beneficial for the model to learn from by interactively querying the user to label the relevant data points.

From the pool of unlabeled data, active learning selects the subset of data to be labeled. The idea is to allow the model to choose the data it wants to learn from and achieve a higher level of accuracy while using a smaller number of training labels.

It is a type of semi-supervised learning wherein a small amount of labeled data, and a large quantity of unlabeled data are used to train a model. The belief behind using semi-supervised learning is that labeling just a small sample of data can result in a similar performance while reducing the enormous effort required to label the entire training data.

Active learning is effective with large amounts of low-cost, unlabelled data, e.g., images and tweets (image borrowed fromhere).

Behind the Scenes of Active Learning

Scenarios: In active learning, there are typically three broad scenarios through which it can be implemented in which the learner will query the labels of instances:

1. Membership Query Synthesis: The learner creates their own instance, in this case, from an underlying natural distribution.

2. Stream-Based Selective Sampling: Here an assumption is made that getting an unlabeled instance is free. In this scenario, each unlabeled data point is examined one by one, the informativeness of each instance is evaluated, and the learner decides whether it wants to query the label of the instance or reject it based on its informativeness. To determine the informativeness of the instance, a query strategy is used.

3. Pool-Based Sampling: Here, the instances are drawn from the entire pool of unlabeled data, and an evaluation of how well the learner understands the data is made by assigning a confidence score or an informativeness measure to each instance, and those instances are chosen in which it is least confident and queries the user to provide the labels.

Query Strategies: Active learning scenarios require some sort of informativeness measure of the unlabeled instances, which is used for determining which data points should be labeled to query the user for them. Each has its own set of pros and cons. These are basically used to assign a priority score to each data point.

1. Uncertainty sampling: An active learner queries the instances about which it is least certain how to label, i.e., the samples which lie close to the classification boundary.

2. Query by committee: Several models are trained on the currently labeled data, and the output for unlabeled data is decided by a vote; the points that the ”committee” disagrees on the most are labeled.

3. Least Confidence: In this strategy, the learner selects the instance it has the least confidence in its most likely label. So, the data points to be labeled are sorted with respect to the confidence in their most-likely labels.

4. Margin Sampling: It selects the instance with the smallest difference between the first and second most probable labels. So, the data points with the lowest difference between the highest probability and the second highest probability would be labeled first.

5. Expected model change: The points that would most significantly change the values of the current model are labeled.

6. Expected error reduction: The data points that reduce the generalization error of the model the most are labeled.

7. Variance reduction: The data points that would minimize output variance are labeled.

8. Entropy: The concept of entropy comes from thermodynamics and is a measure of disorder in a system. The entropy for each instance is calculated, and the instance with the largest value is queried.

9. Density-weighted methods: Points representative of the underlying distribution (that is, reside in the dense areas of the input space) are labeled.

Putting it all together (aka perform active learning)

We have been provided with a large unlabeled dataset for training.

The first thing that must be done is the manual labeling of a very small subset of this data.

Once there is a small amount of labeled data, the model has to be trained on it. The model won't be its best possible version, of course, but it will provide insights into which regions of the parameter space should be tagged first to make it better.

The model is used to predict the class of each of the remaining unlabeled data points after it has been trained. Depending on the prediction confidence of the model, it is decided which data points should be labeled by querying the user for them, or a priority score is assigned to each unlabeled data point.

After the best method for prioritizing the labeling has been selected, the process can be iteratively repeated; that is, a new model can be trained on a new labeled data set that has labels assigned using the priority score. The unlabeled data points can be run through the model to update the prioritization scores and continue labeling after the new model has been trained on the subset of data. In this manner, the labeling method may be continually optimized as the model improves.

A typical example of deep active learning (borrowed fromhere).

Applications

Active learning is widely used in multiple domains and has several applications in the real world. One of the most widely-used applications of active learning is Natural Language Processing (NLP), where the amount of unlabeled data is huge, and the labeling process is tedious. It is used for tasks like Named Entity Recognition, where the model has to find which parts of the input text correspond to entities such as persons, locations, or organizations; Machine Translation; Text categorization, etc., which is the task of assigning a category to a piece of text.

For Machine Translation, curriculum learning frameworks are used that decide which training samples to show to the model during different periods of training based on the estimated difficulty of a sample and the current competence of the model. This method not only effectively improves the training efficiency but also obtains a good accuracy improvement.

Active learning has also proven successful in complex Computer Vision tasks like image segmentation, object detection, image classification, restoration, etc., as a huge amount of unlabeled data is available through the internet. Cost Effective Active Learning (CEAL) merges deep convolutional neural networks into active learning and assigns pseudo-labels to samples with high confidence, and adds them to the highly uncertain sample set queried using the uncertainty-based AL method.

Cost Effective Active Learning mode (image borrowed from theCEAL paper).

Active learning also has applications in the field of speech and audio processing, social network analysis, medical image processing, wildlife protection, industrial robotics, autonomous driving, and disaster analysis, among other fields. Generative Adversarial Active Learning (GAAL) introduced the Generative Adversarial Networks (GAN) used for data augmentation to the active learning query method. GAAL aims to use generative learning to generate samples with more information than the original dataset.

Principle behind the GAAL algorithm. The learner synthesizes samples for querying using GAN.GAAL paper.

Conclusion

Active learning attempts to maximize a model’s performance gain while annotating the fewest samples possible. More specifically, it aims to select the most useful samples from the unlabeled dataset and hand them over to the oracle (e.g., human annotator) for labeling to reduce the cost of labeling as much as possible while maintaining performance. The acquisition of a large number of high-quality annotated datasets consumes a lot of manpower.

Deep learning is greedy for data and requires a large amount of data supply to optimize a massive number of parameters if the model is to learn how to extract high-quality features. In recent years, due to the rapid development of internet technology, we have entered an era of information abundance characterized by massive amounts of available data.

One of the most critical problems faced in the field of image vision tasks is that of efficiently querying samples of high-dimensional data (an area in which traditional methods perform poorly) and obtaining satisfactory performance at the smallest possible labeling cost.

Active Learning is proven to have been successful in many applications and is still an active area of research.