Lex Fridman Podcast #110 – Jitendra Malik: Computer Vision
Jul 21, 2020
Jitendra Malik, a distinguished professor at UC Berkeley and a pioneer in computer vision, shares his insights on the complexities of replicating human visual perception. He discusses the challenges of Tesla's Autopilot, emphasizing the gap between human and computer processing. Malik explores how integrated approaches and knowledge schemas can enhance action recognition. He critiques current evaluation methods, advocating for measures that reflect true understanding. Additionally, he highlights the importance of interdisciplinary research and the need for children's experiences in AI development.
AI Snips
Chapters
Books
Transcript
Episode notes
Misinterpreting a Skateboarder
- Jitendra Malik recounts an example of his Tesla misinterpreting a skateboarder's behavior due to a lack of diverse training data.
- This illustrates the need for AI systems to understand human behavior in various contexts.
Learning to Drive
- Human drivers learn to drive by building upon pre-existing visual knowledge, unlike tabula rasa learning in AI.
- Children develop visual expertise from 0 to 16, focusing on control in driver's ed, not relearning vision.
Vision's Link to Action
- Computer vision's general problem ties to action, like the first multicellular animals needing vision for movement and survival.
- We've evolved general visual capabilities to build internal models of the external world, although imperfect, as shown by illusions.
