One Model To Learn Them All
In an arXiv paper, Lukasz Kaiser et. al. presents a model that's getting closer to artificial general intelligence: " We present a single model that yields good results on a number of problems spanning multiple domains".
https://arxiv.org/abs/1706.05137
Opinions?
Appreciate the work done by the researchers! Designing model architecture has been a hot topic. Building machine learning models is sort of like a repeates experiment - finding appropriate data sources, feature engineering, using different models and again and again. It will for sure be good to have a single model which can produce reasonably good results on a wide range of data set with various data structure to save much time on the work!
Perhaps one day, we can feed machines with all the knowledge we've accumulated, and let it generate more knowledge/innovations that could solve problems we find impossible right now and into the future.
Definitely agree with you!
nice.
Seems like they're finally finding ways to piece together the different parts that allow single specialized AI's to work on certain problems, in such a way a single AI can work on the whole smorgasbord of problems said parts were made for.
I can foresee they'll come up with a follow up article about the AI with the different parts suddenly being able to solve problems they hadn't foreseen.
If they've done it well in the way I infer from their introduction, emergent properties are going to pop up ! :)
Yeah absolutely! Very excited about its applications :D