188 - Guest: Peter Norvig, AI professor/author/researcher, part 1
Jan 22, 2024
auto_awesome
AI professor/author/researcher Peter Norvig discusses the evolution of AI systems, the challenges of AI in chess, advancements and limitations in language processing of computers, and the development of a humanoid robot airplane pilot.
The evolution of AI from the expert system approach to probability-based models and machine learning, as represented by the creation of the textbook 'Artificial Intelligence: A Modern Approach'.
The shift from explicit, rigid symbol-based systems to connectionist models, such as neural networks, which allowed for the representation of concepts as points in a multidimensional space and greatly contributed to recent AI advancements in pattern recognition and learning from examples.
Deep dives
The Development of the Standard AI Textbook
Peter Norvig, co-author of the textbook 'Artificial Intelligence: A Modern Approach,' shares the story behind its creation. In the 1980s, the field was dominated by the expert system approach, where AI systems were built based on self-reports of experts. However, in the late 80s and 90s, there was a shift towards probability-based models and machine learning, which required a new textbook. Norvig, along with Stuart Russell, decided to write a book that represented the changing field. They completed the first edition remotely, and the book has since become the standard AI textbook used in universities worldwide.
The Paradigm Shift in AI towards Connectionism
Norvig discusses the shift in AI from explicit, rigid symbol-based systems to connectionist models. Early AI attempted to define strict concepts and boundaries using logic, but it did not accurately represent the complex, fuzzy nature of real-world concepts. The introduction of connectionist models, such as neural networks, allowed for the representation of concepts as points in a multidimensional space, introducing the idea of nearness and similarity. This shift has contributed to recent advances in AI, enabling neural networks to excel in pattern recognition and learning from examples.
Challenges: The Balance of Pattern Recognition and Reasoning
Norvig acknowledges that while AI has made tremendous progress in language processing, there are limitations that need to be addressed. AI systems often make mistakes and lack the ability for explicit reasoning and explanation. Teaching AI models using explanations and human-like reasoning is still a challenge. The field must find a balance between pattern recognition and formal reasoning to achieve higher accuracy with fewer training examples. Additionally, AI needs to improve its understanding of programming languages and reasoning in various domains. Solving these challenges will enable the integration of human-like reasoning capabilities into AI systems.
Literally writing the book on AI is my guest Peter Norvig, who is coauthor of the standard text, Artificial Intelligence: A Modern Approach, used in 135 countries and 1500+ universities. (The other author, Stuart Russell, was on this show in episodes 86 and 87.) Peter is a Distinguished Education Fellow at Stanford's Human-Centered AI Institute and a researcher at Google. He was head of NASA Ames's Computational Sciences Division and a recipient of NASA's Exceptional Achievement Award in 2001. He has taught at the University of Southern California, Stanford University, and the University of California at Berkeley, from which he received a PhD in 1986 and the distinguished alumni award in 2006.
He’s also the author of the world’s longest palindromic sentence.
In this first part of the interview, we talk about the evolution of AI from the symbolic processing paradigm to the connectionist paradigm, or neural networks, how they layer on each other in humans and AIs, and Peter’s experiences in blending the worlds of academic and business.
All this plus our usual look at today's AI headlines.