As developments in computing power push the AI field deeper and faster, Google is among the tech giants going back to basics and teaching its machines how to talk and read – and using old media to do it
We rely on machines for a mix of everyday and super-human tasks. But other than the odd shouting match with a recalcitrant PC, or asking Apple's Siri to call home, it is rare that humans engage a computer in conversation.
Paradoxically, the limits to machine learning thus far have been due to human constraints. Getting computers to talk previously required humans to code rules governing each interaction, so the computer would know how to respond.
Now, Google thinks it can get machines to answer questions on their own by learning from old movie dialogue. In a paper released in June, two Google researchers tested their approach by getting computers to 'talk' based on what they learnt from film subtitles. Using artificial neural networks - where large numbers of computers connect and adapt to inputs - machines could predict what the next sentence should be, given the previous one. The resulting model 'is sometimes able to produce natural conversations', said the paper.
Another Google-backed outfit is teaching its computers to read. London-based artificial intelligence (AI) company Google DeepMind uses online articles from news outlets The Daily Mail and CNN as a learning database. What makes these stories perfect for machines is the way they are set out: with summary bullet points that precede the main story. Crucially, the bullet points use different phrases from the main text, creating annotations from which machines can learn grammatical links and causal relationships, enabling it to answer questions based on what it reads. It is only a matter of time before AI catapults computers to the top of the class.
www.roboticsforgood.ae
