Transfer learning reuses existing AI algorithms for new tasks
” src=”https://s.yimg.com/ny/api/res/1.2/q7Y.P8enjX09F6fpraan_w–/YXBwaWQ9aGlnaGxhbmRlcjt3PTk2MDtoPTY0MA–/https://media.zenfs.com/en/lifewire_488/d72f15d2eff7fa1ca6a050e6b16bb425″ class=”caas-img”/>
In machine learning, transfer learning occurs when an existing algorithm takes on a different (but similar) job. Here, we provide an overview of transfer learning, its benefits, and applications.
Definition of transfer learning
In transfer learning, developers reuse algorithms designed for a specific purpose for different tasks. New algorithms apply what they already know to do new work.
For example, if you have an algorithm that can identify pictures of dogs, you can easily adapt it to identify cats as well. You can even build an algorithm to create an algorithm that can identify any animal.
Artificial intelligence (AI) programs rely on various machine learning algorithms to perform their intended tasks better and faster. Transfer learning is not really a form of machine learning, but rather a technique used in the field. Transfer learning has applications other than machine learning.
Related: Machine Learning and Deep Learning: What’s the Difference?
Imaginima / istock / Getty Images
” src=”https://s.yimg.com/ny/api/res/1.2/oVeZe9LCgAxdbaXzqvbWzA–/YXBwaWQ9aGlnaGxhbmRlcjt3PTk2MDtoPTY0MQ–/https://media.zenfs.com/en/lifewire_488/e4bad495af6c00b616e24ad726dc4050″ class=”caas-img”/>
Imaginima/istock/Getty Images
Applications of transfer learning
One of the most attractive prospects for transfer learning is deep learning. Using transfer learning as part of a neural network, AI can be trained how to solve new problems, even with very little data available.
Real-world situations can’t always be decomposed into data points, so a strong AI must be able to make inferences based on past experience. Transfer learning has brought the promise of artificial general intelligence (AGI), an AI that can perform any task that humans can do, closer to reality.
Advantages of transfer learning
Transfer learning is more attractive than creating a new algorithm from scratch. Training a neural network can require a lot of time, data, and computing power. In addition to saving time, building on pre-trained models can improve results.
Thanks to transfer learning, AI programs can be trained to perform tasks that otherwise would not be possible. If you don’t have enough data to train a neural network for your desired job, you can train the network to perform a similar task with plenty of data. You can then build on that model and successfully train the network for new tasks using limited data.
Types of transfer learning models
There are dozens of pre-trained algorithms available for AI engineers to build. You can use only part of an existing model, or you can use the entire model. Alternatively, you can build your own algorithm and reuse it.
Image classification, object recognition, and computer vision are common applications of transfer learning. Transfer learning models used for image recognition include Google Incept and Microsoft ResNet. These models are open source and available for anyone to use.
Another promising application of transfer learning is natural language processing, especially text-to-speech translation (or vice versa). Google’s Word2vec and Stanford’s GloVe are two public models that can be adapted for deep learning language projects.
The Caffe Model Zoo is a community-run website with a repository of pre-trained models for transfer learning projects and tutorials on how to use them.
FAQ
What is transfer learning in CNN?
Transfer learning in convolutional neural networks (CNNs) is the same process as in other areas of machine learning. Existing algorithms migrate to apply their “knowledge” to new tasks. CNNs may differ in structure from other machine learning systems, but the process works in the same way.
When should I use transfer learning?
You will need to use transfer learning at various points in your machine learning project. One, you’re confident that you’ve “learned” everything your current model can do, and you want to extend its capabilities. Another valid use is when the scope or purpose of a project changes. Transfer learning helps you pivot without completely starting over.
