![]() |
Engadget |
Last week, they published data sets which showed how they train their big AI neural networks. It was all fairly standard, binary laden fare, but one thing in particular stood out: they were using children's books. You can't program an understanding of language directly into an AI system, it needs to be taught, bit by bit, and what better way of doing that than children's literature?
One of the data sets released was a set of formatted children's books, including Alice in Wonderland, The Jungle Book, A Christmas Carol and Peter Pan. There are around 100 stories total, all taken from a library called Project Gutenberg and reformatted so that the system can more easily read and process them. Once read, the system is tested on the stories, but, in an interesting twist, they were tested on the books which they hadn't read.
![]() |
An example of one of the tests (via New Scientist) |
The same AI has also been tested on things like general knowledge and relationships between objects in stories, all of it designed to make the system more situationally versatile. You can see how that kind of development could apply to things like Siri, but Facebook has far higher ambitions for it, let's just hope they approach them with some measure of caution.
Post a comment