Here is a paper that uses meta-learning for quick imitation learning in robots that works across domains:
<a href="https://arxiv.org/abs/1802.01557" rel="nofollow">https://arxiv.org/abs/1802.01557</a><p>The main sell is that it allows the robot to quickly generalize to new domains (surroundings and objects being manipulated), albeit not tasks.
I enjoy taking ML work and pondering what it might look like in the context of human education.<p>Imagine human physics education with meta-learning characteristics. Having to classify problems. No more "here are numbers for <i>solid</i> Argon... apply the <i>ideal gas</i> law" clueless plug-and-chug. Developing a sense for reasonable values. Encountering unfamiliar problems. Thus developing skills of rough quantitative reasoning, and of system decomposition and characterization. Encountering descriptions of unfamiliar problem domains. And having to extract understanding from them.<p>With human science education at present, even correct labeling and baseline models, let alone transference, are distant, distant dreams. But even thus shackled and buried, visualizing dance might have value, if it supports improved recognition of opportunities and preparation for escape.
I was disappointed the article was about ML and not about human learning.<p>When I was young, I got by on picking basics up quickly. Now I'm in my forties and I'm having to actually learn how to learn because there are no simple basics (in my areas) left (or I'm just old and the basics no longer seem so basic). Everything I read/listen/watch is a collection of boring rehash...and then I'm suddenly behind because I missed something while skimming the material I thought I knew.<p>The article title got me excited. :(