Right now, AIs have training and inference phases. Humans also have awake and sleep phases. But humans can learn, validate and immediately use stuff what we learn while awake. Is it possible to give AI this “inference with learning” superpower?
Inference that allow to learn directly from the user, retain and enhance on that knowledge, while being able to validate. Then it is possible to compress the “domains” into faster structures. This way model evolves its weights efficiently without spending too much resources. Like a person in a bar would do. Not like a person finetuned in a university.
While sleeping has more to do with maintenance of the immune system than learning (w/o sleep even fruit flies die and night work is correlated with higher risk of cancer), the answer is that nobody knows how do “one-shot learning” in machines. We are still in the times of alchemy in that respect.
The one approach that tried to develop a general model for learning (symbolic AI trying to emulate what behavioral and psychological studies reveal about learning) was Newell/Laird/Rosenbloom’s Soar but it is talked about rarely nowadays.
That being said, I have just read these old-ish articles they might be of interest regarding the questions raised:
December 2004: Moving AI out of its infancy: changing our preconceptions
May 2007: In the News: Bridging the Gap between Neuroscience and AI
1 Like
Hi @abitrolly ,
The question of training AI by analogy with human learning is addressed in an noteworthy article by F. Chollet On the Measure of Intelligence. He is also the founder of a competition aimed at developing AI systems that can learn similarly to humans, acquiring new skills from a small number of examples - ARC Prize Competition.
1 Like
Interesting but that paper seems to be about trying to establish a benchmark for intelligent system, it is not about learning 
Using algorithmic information theory … does that work? One hits the “uncomputable” skids pretty much immediately in that branch.