Active machine learning (active ML) is a powerful approach that seeks to create predictive models with remarkable accuracy, all while minimizing the number of labeled training examples required. This is achieved by employing a clever strategy that involves selectively choosing the most informative data points to be labeled by a knowledgeable oracle, such as a human annotator. By doing so, active learning enables models to extract the necessary knowledge they need from a relatively small amount of data.

Now, let’s explore some definitions and the fundamental concepts that form the foundation of active ML.

Definition

Active learning can be defined as a dynamic and iterative approach to machine learning, where the algorithm intelligently engages with an oracle to label new data points. An oracle is a source that provides labels for data points queried by the active learner. The oracle acts as a teacher, guiding the model by...

Active learning can be implemented in different ways, depending on the nature of the unlabeled data and how the queries are performed. There are three main scenarios to consider when implementing active learning:

• Membership query synthesis
• Stream-based selective sampling
• Pool-based sampling

These scenarios offer different ways to optimize and improve the active learning process. Understanding these scenarios can help you make informed decisions and choose the most suitable approach for your specific needs. In this section, we will explore each of these scenarios.

Membership query synthesis

In membership query synthesis, the active learner has the ability to create its own unlabeled data points in order to improve its training. This is done by generating new data points from scratch and then requesting the oracle for labels, as depicted in Figure 1.2. By incorporating these newly labeled data points into its training...

In traditional passive machine learning, models are trained on fixed and pre-existing labeled datasets, which are carefully assembled to include both data points and their respective ground truth labels. The model then goes through the dataset once, without any iteration or interaction, and learns the patterns and relationships between the features and labels. This is the passive learning approach. It’s important to note that the model only trains on the finite data it is provided and cannot actively seek out new information or modify its training based on new inputs. Moreover, the labeled datasets required for a passive learning approach come at a cost.

There are several reasons why labeling is expensive in traditional machine learning:

• Manual labeling requires experts: Accurately labeling data often demands the expertise of domain specialists such as doctors or ecologists. However, their time is limited and valuable, making...

In this introductory chapter, we covered the fundamentals of active ML and how it contrasts with passive learning approaches.

You learned what active learning is and its goal of maximizing predictive performance with fewer labeled training examples. We discussed the core components of an active learning system: the unlabeled data pool, query strategy, machine learning model, and the oracle labeler.

You now understand the difference between membership query synthesis, stream-based sampling, and pool-based sampling scenarios. We compared active and passive learning, highlighting the benefits of an interactive, iterative approach in active learning.

Importantly, you now know that active learning can produce models with equal or greater accuracy while requiring far less labeled training data. This is critical for reducing the costs of modeling, as labeling is often the most expensive component.

The skills you gained in this introduction will equip you to determine when...