The engineers at the company DeepMind created a machine-learning platform that was based on research into the brain function of babies and how it performed in certain tasks than its counterparts in the traditional.
Hristopher Intagliata Artificial Intelligence systems have beaten humans in poker, chess Jeopardy, Go, and numerous different games. However, they aren’t proficient at understanding fundamental rules that govern how the world works.
Susan Hespos: It’s still not possible for them to perform the same things that toddlers do. And I’m a fervent advocate for babies at the end the day. This is an obvious win for infants. Babies continue to be the most advanced computers in the ability to comprehend physics in a way that is intuitive.
Intagliata Psychologist Susan Hespos from Northwestern University listed off a couple of examples of these “intuitive physical” principles. As an example, there is “solidity” which means that your coffee cup doesn’t slide right through the table. Also “continuity” objects don’t disappear and come back into existence. Also “boundedness” is when you grab your cup of coffee and it sticks together. It doesn’t leave you only with the handle.
Hespos Baby’s can comprehend all of the three from the time they are three months old. Their visual acuity is poor and their world is blurry…they are unable to comprehend this information. Babies can do lots of mistakes. However, it’s the initial kernels that become more elaborate and refined by experience across the globe.
Intagliata Computer engineers are now taking a leaf out of the playbook of babies. Researchers from DeepMind – the AI company which developed computers that beat humans in the game of Go in the past — infused an artificial intelligence system with specific kernels information about physics that is intuitive into… like what a newborn could have.
Then, after having watched what amounts to just 28 hours of videos for training, that showed things like balls rolling and blocks falling -the AI system displayed “surprise” when shown something that is physically impossible. The AI system’s peers which were not modelled on infants weren’t as precise.
Hespos This is really fascinating because when you make this kind of direct comparison, what you discover from the experience can be quite extensive. However, it’s only up to a certain point. And the computer designed based on research into babies, performed much better. It’s providing proof of the research on babies that has been conducted for some time, but using something completely different from the infant.
Intagliata: The findings appear within the journal Nature Human Behavior. (Luis S. Piloto et al., Intuitive physics-based learning in a deep-learning model influenced by psychological development[Luis S. Piloto et al, Intuitive
Hespos did not participate in the research She did write an editorial to accompany the research paper. She claims that the study will help make machines more effective thinking machines, just like humans. Even the smallest ones.