It’s no secret that the goal of artificial intelligence (AI) is to teach machines to think and act as humans do. That may seem like an impossible task, but we’ve seen remarkable achievements in the field. The primitive human act of communication, for instance, has been taught to machines in an area called “conversational AI.”
And would you believe that programs can detect a person’s emotions and respond accordingly? Experts are studying artificial emotional intelligence extensively in what we know as “affective computing.”
Given the variety of fields that AI touches, it’s no surprise that experts are also exploring how humans learn. Transfer learning is an example, and in this post, we’re going to learn what transfer learning is and how it works.
What is Transfer Learning?
Transfer learning is an AI technique where the knowledge gained by machine learning (ML) models in solving one problem gets applied to another.
Humans find transfer learning easy. We automatically apply what we learned in previous experiences when we face a new task. Consider this: Do you recall being told that once you know how to ride a bicycle, learning how to ride a motorcycle would be a piece of cake?
In task 1 (learning to ride a bike), you learn how to balance, and once you learn that, you don’t have to start from scratch when you do task 2 (learning to ride a motorbike). That’s because of transfer learning, which is innate in humans. Machines, on the other hand, need to be taught to learn the same way.
How does Transfer Learning Work?
If you train a machine to detect the presence of people in images, you can use that knowledge to make it recognize who the person in the picture is. Take note that both tasks make use of the same input—images.
Also, they have different but related outputs. Task 1 aims to detect the presence of people in images. Task 2, meanwhile, requires identification of the people in the photos, much like how the tagging feature in social media platforms works.
The machine’s knowledge from task 1 can help accomplish task 2, making it easier and more cost-effective. Since it already knows how to detect people in images, you don’t have to start from scratch to teach it to do task 2.
In essence, transfer learning is the act of reusing a model on a new problem or task. Now, transfer learning can be done in part or in full. Everything from the model is reusable. But there may be instances when only certain features of the model can be reused.
Either way, training becomes faster, and the results, more reliable. Check out IBM Watson’s short video below. It talks about how the vendor uses transfer learning to speed up insurance claims processing by 25%. In the process, users can significantly save on operating costs.
What are Some Applications of Transfer Learning?
In the video above, you learned how transfer learning helps home insurance companies. The learning technique can also work in other sectors, such as:
1. Image Categorization
As illustrated in the previous section, transfer learning helps detect and recognize objects in images. As a result, it makes image categorization a lot easier and faster. You no longer have to train a model to classify images since it’s already pre-trained.
The process is beneficial in several sectors, most notably in the medical field. Medical imaging apps no longer have to be trained from scratch and can immediately detect abnormalities such as kidney stones in scans.
The application of AI in gaming has been explored as early as the 1990s. Different machines have defeated human game masters. We’ve seen IBM’s Deep Blue for chess, Chinook for checkers, and AlphaGo for the classic board game “Go.” These inventions, however, would not be able to win other matches. For example, if you use DeepMind’s AlphaGo to play chess, it would fail because it uses the traditional method of learning.
With the help of transfer learning, though, Deepmind developed another system that can use the strategies it learned from playing “Go” to play a different game.
Although machines can’t fully replicate the way humans learn, mimicking even a portion of it can still change the AI landscape. Given enough datasets, transfer learning can cut down the costs and time in developing intelligent machines. Developers can reuse ML models in related tasks or problems instead of starting from scratch.