Ai’s researchers developed a computer program designed to learn how to carry a conversation and succeeded in training it to converse on the level of a 15 month old baby.
Using Alan Turing’s Child Machine test and a developmental model, Ai’s research division created Hal, Ai’s baby algorithm, which is said to demonstrate the linguistic capabilities of a 15-month-old baby.
According to Ai, this is just one of a series of milestones, based on human development, that Hal is expected to pass before it achieves the conversational capabilities of an adult speaker.
Previous attempts to develop conversational computer programs are said to have used crude techniques when replying to questions or statements.
The program would concentrate on a key word, and then use statistical techniques and a formal understanding of grammar to generate appropriate replies or pick them from a pre-generated list.
Hal’s algorithm is built using general learning principles developed by Ai’s algorithm department and is then trained by a team of cognitive scientists and child development experts to acquire language.
The learning algorithms underpinning the program gradually learn what are appropriate responses and how to react to the conversational style of its tutor.
The first milestone was achieved following combined efforts by both departments within Ai’s Research Division. Hal’s conversations were presented to external language experts who could not distinguish between the lingual performance of the algorithm and that of a 15-month old baby.
‘This is an extremely exciting development, bearing in mind that HAL was not programmed with any pre-conception or assumption about language or the world around it,’ said Dr Anat Treister-Goren, Ai’s Chief Trainer. ‘Its performance is strictly the result of the application of general learning principles to behaviourist training methodologies.’
Ai’s Research Division is now working on its next challenge – attaining the linguistic performance of an 18-month-old child.