What is transfer learning and why it is so important to artificial intelligence
"In essence, transfer learning is about speeding up a new learning task by reusing the results of previous learning." Deep Learning with JavaScript
One of the most astonishing details about learning is their compact form. Behind E=mc², we have a sea of scientists. Its simplicity has nothing to do with its size, but its ability to grasp so shortly centuries of science progress. With one set of symbols in you hand, you just explained how the atomic bomb is created. I do not want to even get started on Schrodinger equation.
Would it be possible to replicate this pattern in machine learning?
In part, yes. If you mean, like humans, not. This ambition is strongly inside artificial general intelligence.
As one example, during my postdoc, where I came to know about transfer learning for the very first time, they used transfer learning to train medical image segmentation.
"Transfer learning sometimes involves retraining the base model and sometimes involves creating a new model on top of the base model." Deep Learning with JavaScript
Detail: the main training dataset was done using normal objects, unrelated to the medical situations! See my report online. For me, when I first saw the results, I could not believe, and still cannot. Nonetheless, since I am not new to neural networks, I know they work like magic, and we must accept our ignorance.
Thus, transfer learning is an approach where one can reuse the learning process from one scenario to other: this is different to use a pre-trained model for prediction, it is more like get someone already graduated, and make some final adjustments to become a nice professional.
It seems the closest we got from the human ability to abstract concepts aways. See that this is not an abstract in the human sense, it is still the old narrow learning machine. Even though it seems we have abstracted as humans, there is no evidence we actually did that.