Are artificial intelligence learnings or trainings transferable from one agent to the other?

Question

One disadvantage or weakness of Artificial Intelligence today the slow nature of learning or training success. For instance, an AI agent might require a 100,000 samples or more to reach an appreciable level of performance with a specific task. But this is unlike humans who are able to learn very quickly with a minimum number of samples. Humans are also able to teach one another, or in other words, transfer knowledge acquired.

My question is this: are Artificial Intelligence learnings or trainings transferable from one agent to the other? If yes, how? If no, why?

Answer 1

The simplest case is copying the software. That is instant duplication of the learnings. In a similar, but less trivial way you can adjust pretrained neural network classifiers to new datasets by simply re-initializing the last layer.

It gets more interesting when you have multiple agents and you want to combine their knowledge. This can be done with many techniques; the simplest is averaging the weights of identical neural network architectures. One example of this combined learning happened is Google Has a Room Full of Robot Arms Learning Hand-Eye Coordination. The relevant paper is Learning Hand-Eye Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection.

The most complicated case, where you want to transfer just the gist and prevent forgetting is -- up to my knowledge, and I'm relatively certain it's outdated in that area -- unsolved. The problem is post-hoc explanations. Humans decide to do something, observe the outcome and then explain why they did something / why it was good to do it. So we automatically generate hypotheses about the world. And sometimes we are able to formulate them in a way that others can understand them. That is not possible with current learning machines. We don't know yet how to abstract automatically from arbitrary problems.