Microsoft Reseach’s experimental AI is already a leader in machine reading, being able to read a piece of text and synthesise succinct answers from the text, allowing a search engine, for example, to answer a question rather than simply deliver a page which may or may not contains the answer.
The approach currently used however relies on an AI being trained with manually labelled questions and answers which is not scalable. Now in their latest move, Microsoft has created an AI which can be trained on one set of labelled text to discern what items contain interesting data, and then automatically generate question and answer pairs from the data. See an example below.
The AI is then able to use this skill to generate labelled question and answer pairs from a new set of documents and data, which can then be used to train a second AI to become an expert in that field.
Using the SynNet system, Microsoft has been able to get more accurate results on a new domain without any additional training data, approaching to the performance of a fully supervised MRC system.
Readers who have heard Elon Musk’s recent warnings may recognise that, by learning how to teach other AI, this means AI have a much easier time generating knowledge from all the information scattered around the internet. The less paranoid may read Microsoft Research’s full article for more detail on the process, which Microsoft, of course, hopes to “rapidly expand” in the future.