Smash Notes
UC Berkeley's Alison Gopnik: "Babies are the ultimate supercomputers"
The Sunday Times’ tech correspondent brings on Dr Alison Gopnik, a developmental psychologist, to talk about why babies could be the key to artificial intelligence (3:45), the limits to current systems (5:40), infants as supercomputers (8:00), the power of experimentation (10:15), how young brains learn (12:50), coding curiosity (16:15), how the tech…
People
Notes
-
Techniques that have been used aren't powerful enough to do the type of learning that humans do. The people who do most of that learning are babies and young children.
-
From small amounts of data, they can make very impressive generalizations. Studies show that they are often better than adults at coming up with unlikely ideas.
-
Alison Gopnik tells Danny Fortson that children solve problems differently from current AI. Her acronym is MESS: model building, exploration, and social learning. Model building is the claim of her book The Scientist in the Crib. Children hold theories, test them, and make counterfactuals, more than spotting correlations. Exploration is what adults call getting into everything, an explore-versus-exploit tradeoff. She cites the Goldilocks effect, work by Berkeley colleague Celeste Kidd: even a nine-month-old looks longest at a just-right amount of new information. Social learning is watching people nearby and trying to join in. She sets AlphaZero, which can beat chess grandmasters, against her grandson Addie throwing the pieces in a wastebasket and inventing a new game.
-
On Danny in the Valley, Alison Gopnik of UC Berkeley describes work with Berkeley computer scientists, including Pulkit Agrawal and Deepak Pathak. They call it curiosity-based reinforcement learning, drawn from developmental psychology. Ordinary reinforcement learning tells the agent that a score went up or down. Here, the reward is for a prediction that does not fit what you already know. When something unexpected happens, you go back and try to make sense of it. Gopnik says that looks like what children do. Her lab was giving the same problems to four-year-olds and to the curiosity-based AI, to compare what each one did.
-
Babies and young children are born with a lot of ideas of how the world works. Children are also actually doing the most learning; a child's brain is going to be better at learning than an adult's brain.
-
Part of it may involve determining if it is best to develop systems that humans can't do well vs. what they can do well (i.e. processing enormous amounts of data vs. creativity/curiosity).
-
Yes. The general idea of the way to go is to continue to develop these structured models surrounding the brain development of babies and young children.
Related notes
- How Do Raia Hadsell's Colleagues Catch a False Win?
- Are the Scientists Trying to Make Tools to Solve What Humans Can’t Solve or Replicate Human Intelligence
- How Do You Program an AI to Produce Something That Is Pure Creative
- Why Does a Researcher Join an Artificial Intelligence Company?
- What Areas Are We Seeing AI Applied in Our Normal Lives