Tomaso Poggio: Brains, Minds, and Machines | Lex Fridman Podcast #13

Poggio points out that central techniques like deep learning and reinforcement learning were inspired by neuroscience and behavioral science (Hubel & Wiesel, Pavlov, Minsky). He bets that at least some of the next major advances in AI will again come from understanding brain circuits and learning mechanisms.

When problems can be decomposed into hierarchies of simpler sub-problems (vision, language), deep architectures can avoid the curse of dimensionality that plagues shallow models. Poggio argues that either physics is compositional—or our brains are wired as deep networks and we only experience and care about compositional problems we can solve.

Modern deep nets often have far more parameters than data points, yet train successfully. Poggio notes that this creates an enormous number of global minima—“probably more minima than atoms in the universe”—so simple algorithms like stochastic gradient descent are surprisingly likely to find good solutions; the harder question is why some of those solutions generalize well.

Where current AI needs vast labeled datasets, children learn from very few explicit labels. Poggio suggests evolution supplies weak biases (e.g., motion sensitivity, quickly imprinting common stimuli like faces), and then experience bootstraps richer representations, a process today’s supervised methods only dimly resemble.

Despite distinct functions (vision, audition, language), cortical regions share remarkably similar microcircuitry. Poggio sees this as evidence for a common underlying architecture that, via learning and connectivity, gets specialized—an important clue for designing more general AI systems.