Skip to content

News & Insights

Graphic of a circular ring

I is for Incremental Learning


I is for Incremental Learning

From A to I to Z: Jaid’s Guide to Artificial Intelligence

Incremental learning is a machine learning method in which AI is trained in stages. 

The process broadly mimics how students learn about a subject in school. Just as you’d start with the basics and broaden your knowledge with every subsequent lesson, incremental learning enables AI to acquire more information and master skills over time. 

Incremental learning has three key benefits. 

First, it avoids the risk of what is known as catastrophic forgetting. This is a phenomenon in which the AI forgets what it previously learned when fed a huge amount of new data. 

Second, it’s more efficient. AI needs far less time and computing power to process data if it’s fed to it in batches, instead of all at once. 

Third, it enables the AI to continuously improve. Should the data change or new data become available, you can feed this to the AI instead of having to restart training on the whole dataset from scratch. 

Incremental learning is most commonly used in cases where AI must process constantly changing data in real time. Examples include:

  • Recommendation engines like those used by Spotify and Netflix
  • Stock market algorithms and other software that analyzes data with constantly shifting patterns
  • Medical diagnostics
  • Autonomous systems, like self-driving cars, that need to adapt to changing circumstances
  • Online learning systems that have to adapt to different students’ levels and abilities

Some facts:

Researchers began exploring incremental learning in the late 1980s. Back then, the prevailing view was that, if you trained a computer on one “good” data set, it would be able to correctly apply that knowledge to new scenarios. So incremental learning was a huge departure from conventional thinking. 

In a 1986 paper, computer scientists Jeffrey C. Schlimmer and Douglas Fisher challenged this view, observing that “Learning effectiveness in complex domains requires the development of incremental, cost-effective methods.

Needless to say, they were proved right, and incremental learning has become one of the cornerstones of modern machine learning methodology.

Incremental learning has opened up all sorts of new possibilities. The technique has since been used to teach AI a broad range of skills, including how to play video games, write short stories, and compose music.

There are four main types of incremental learning:

  • Online learning, where the AI is fed data points in real time, as and when they become available
  • Active learning, where the AI chooses which data to train on by itself
  • Transfer learning. Here, the AI is trained to master a specific task using knowledge it gained from a previous task
  • Multi-task learning. As the name suggests, this method entails training to do several related tasks at the same time

Want to know more?

This is the groundbreaking 1986 paper in which Schlimmer and Fischer make the case for shifting to incremental learning techniques. Notably, the paper concludes that “As machine learning methods are applied in more complicated domains, the deficiencies of nonincremental, search intensive methods have become evident.” Today, this seems like a statement of the obvious, but it was somewhat controversial at the time. 

While incremental learning has many advantages, it also has its challenges. This video discusses some common issues with the technique and ways to get around them.

Jaid’s perspective

Incremental learning is a powerful training technique, because it enables AI to continuously broaden its knowledge and get better at what it does. In a customer service setting, this means AI can draw on previous interactions to make future ones more accurate and satisfying for customers.

Take advantage of Jaid’s AI-powered platform to optimize your business functions – contact us today!