What is Transfer Learning?

Transfer learning is a concept in Artificial Intelligence.

Suppose you sit at a neighbourhood kirana store. You notice that the dal that is sold the most is Tur/Arhar dal, and the kirana shop sells about 10 kilos of this dal every day. On Saturday and Sunday, twice this amount is sold because many families buy ration on the weekend. Based on this analysis, after 20 days, you are able to advise the kirana store that they should order maybe 100 kilos of arhar dal per week. This way, they will not run out.

What has just happened is that you have learnt from the data available to you. You have then made a prediction based on that information.

This is predictive AI in its simplest form.

Now, suppose, next month, your friend offers to sit at the kirana store.

She observes the kirana store for 20 days too, and arrives at the prediction that the store should order about 20 kilos of sugar per week.

Working independently, you can predict the weekly order quantity of tur dal and she can predict the weekly order quantity of sugar.

But, lets consider a third story.

In this story, you tell your friend when she starts her observation, what you have found about the tur dal quantity.

Now, at the end of her twenty days, she will be able to:

A. Predict the quantity of tur dal

B. Predict the quantity of sugar

C. Check and predict how many customers are likely to order BOTH tur dal and sugar.

The act of one AI (artificial intelligence) model learning from an earlier model is called transfer learning.

Where is transfer learning used?

Absolutely anywhere that AI models need to learn. Much of our existing learning goes into training future AI models. It is used in Chat GPT, AI image creators, making deepfakes – the applications are literally endless.

Image credit: The image of 2 friends has been generated using AI (Bing). Featured image is from Pixabay.