AI learns like a real toddler

An experiment to create a realistic artificial intelligence consists of a virtual child that interacts with humans and learns in real-time.

babyx.jpg
Auckland Bioengineering Institute Laboratory for Animate Technologies

When it comes to learning, there's nothing quite like the mind of a young child. When they are born, their brain is still developing, and continues to do so for years to come. Although the human brain never stops changing throughout our lifetimes, in those early formative years, it's basically a machine for soaking up information and experiences.

For this reason -- and because brain activity is famously hard to recreate artificially -- it might just be the perfect starting point for AI. So the toddler brain is the perfect subject for an experiment by a team led by Mark Sagar, director at the Laboratory for Animate Technologies at the Auckland Bioengineering Institute and former Weta Digital special projects supervisor.

Called BabyX, the toddler simulation can mimic facial expressions, read basic words, recognise basic symbols, and respond to positive feedback. She also gets distracted, much like a real child, losing her focus to stare at the camera in the room.

"BabyX is an interactive animated virtual infant prototype," the description reads on the Auckland Bioengineering Institute website. "BabyX is a computer generated psychobiological simulation under development in the Laboratory of Animate Technologies and is an experimental vehicle incorporating computational models of basic neural systems involved in interactive behaviour and learning."

BabyX incorporates a number of algorithms, mimicking both neural behaviour when it comes to learning, and responses to the world around her in real-time -- such as the effect of dopamine when she receives praise, for instance. She can also see and hear, and her model employs realistic facial simulation. The researchers use biologically plausible learning models, such as repetition, association, conditioning, imitation, and reinforcement learning, to teach BabyX basic skills.

The project is accompanied by interactive real-time models of BabyX's neural systems and neural anatomy so that the researchers can see the internal processes generated by the computational simulations.

"BabyX embodies many of the technologies we work on in the Laboratory and is under continuous development, in its neural models, sensing systems, and also the realism of its real time computer graphics," the project page reads.

Check her out in the videos below.

 

Join the discussion

Conversation powered by Livefyre

Show Comments Hide Comments