Interview
Gary Marcus: Nature vs Nurture is a False Dichotomy | AI Podcast Clips
- The speaker rejects the "nature versus nurture" dichotomy, asserting that innate biological structures and learning mechanisms must interact; one cannot exist without the other.
- A common misconception in machine learning fields is the false belief that focusing on innate mechanisms constitutes "cheating" against learning models.
- Innate knowledge is traced to the biological development of the brain from stem cells through the central nervous system, encoding information accumulated over evolutionary timescales.
- The speaker cites a baby ibex climbing a mountain shortly after birth as evidence of innate, unconscious knowledge regarding body physics and 3D geometry.
- Developmental psychology suggests humans are born with foundational frameworks for space, time, other agents, places, and mental algebraic reasoning regarding causation.
- These cognitive mechanisms are not disjoint but interrelated; systems representing space and time must communicate to reason about causal events involving physical objects.
- Evolution is described as a "terribly inefficient" process regarding the initial discovery of concepts but highly efficient at spreading and refining those ideas once a viable solution emerges.
- It took approximately one billion years to evolve the vertebrate brain plan, which subsequently spread rapidly across diverse species like fish, dogs, and primates as a reusable biological "library."
- Similar to software subroutines, evolution reuses genetic "libraries" to create specialized adaptations (e.g., using the same digit-building logic for hands and feet) with slightly different parameters.
- Current attempts at evolutionary computation often fail due to starting with minimal innate structures, causing slow progress and making it difficult to generate successful innovations.
- Cumulative evolution is critical for accelerating intelligence; unlike random independent trials (the "monkeys on typewriters" analogy), cumulative processes allow fitness functions to drive rapid optimization.
- The speaker argues that while biological evolution is not logically required to build intelligent systems, human engineers can accelerate the process by borrowing from biological and cognitive science.
- The speaker advocates for "biomimicry" in AI, specifically looking to cognitive science, neuroscience, and linguistics to understand how creatures like dogs intuitively reason about physics and other agents.
- Understanding "dognition" (dog cognition) and its implementation could provide direct engineering insights for creating more advanced AI systems.
- The speaker proposes that the most effective way to bypass the long timescales of biological evolution in AI development is to study and mimic the problem-solving strategies already established in nature.