Tomaso Poggio — MIT professor and director of the Center for Brains, Minds, and Machines; foundational researcher in computational neuroscience and the science of intelligence, who advised Demis Hassabis, Amnon Shashua, and Christof Koch.
The gist
Lex Fridman talks with Tomaso Poggio about why the problem of intelligence may be the greatest problem in science. They discuss how neuroscience has driven recent AI breakthroughs like deep learning and reinforcement learning, the differences between biological and artificial neural networks, and the curse of dimensionality that compositional deep networks help overcome. Poggio examines how children learn from few examples versus data-hungry supervised learning, the role of weak evolutionary priors, and whether brain modules like the face area are innate or quickly learned. The conversation ranges into consciousness, ethics, existential AI risk, and what it takes to do great science.
Big reveals
- Poggio argues recent AI breakthroughs (deep learning, reinforcement learning) originated from neuroscience and bets future breakthroughs will too.
- His hunch is the brain's face-recognition area is not hardwired but a plastic region designed to be imprinted very quickly in early life.
- He partly disagrees with Max Tegmark: rather than compositionality coming purely from physics, our brains may be wired as deep networks so we only solve compositional problems.
- Deep networks have more good (zero-loss) minima than there are atoms in the universe, due to over-parameterization, explaining why optimization works.
- Poggio's proof: deep hierarchical networks with local connectivity can completely avoid the curse of dimensionality for compositional functions.
- He calls claims that AI is more dangerous than nuclear weapons (by Musk and Bostrom) misleading, saying nuclear weapons deserve more worry.
- Researchers can stimulate specific brain areas with magnetic fields and actually change a person's ethical decisions.
Things worth remembering
- Einstein was the worst of five physics PhD students at ETH Zurich and the only one who didn't get an academic position after graduating.
- Human DNA does not have many more genes than the fruit fly Drosophila, which is roughly 95 percent hard-coded by genes.
- Marge Livingstone raised baby monkeys deprived of seeing faces; their brains developed no face-preference area.
- In that deprivation experiment, monkeys often saw the technicians' blue gloves, and some brain cells became hand-sensitive instead of face-sensitive.
- In 1950s neuroscience it was briefly believed the brain was equipotential, that any part was equivalent to any other.
- Reading works compositionally: you don't read every letter but combine syllables into words and words into sentences.
- The universal approximation theorem is essentially the same as a much older result that any continuous function can be approximated by a polynomial.
- In a Google X experiment, subjects seeing through a nearby robot's cameras felt their self had moved to where the robot was.