How can l explain why transfer learning gives better performance than standard training?
August 21, 2017 / Ask Slater
If you’re looking for an analogy that will explain the concept to a non technical user (unclear from question) here are my two favorite analogies:
Imagine you were trying to learn sentiment analysis in Mandarin. Eventually you would be able to figure it out if we showed you enough examples, but it would take a long time because you don’t have a base level of information helpful for the task (speaking Mandarin in this case). Transfer learning works under the same principles. It involves giving your models a base level of information, not specifically related to the task at hand, that still improves your ability to perform the new task.
On the image side I try to explain with the example of teaching an infant what a tiger is. Since an infant doesn’t know anything about the world, they will eventually figure it out, but it will take a long time, and require more examples. Compare this to teaching a toddle what a tiger is. They might not know what a tiger is, but you can tell them that a tiger is a big orange cat with stripes, and because they understand those basic concepts (big, orange, cat, stripes), they can learn the new object (a tiger) more quickly.