Microsoft and Numenta Explore True Machine Intelligence

A Dialogue on the Future of AI

As part of the series The Shape of Things to Come, Doug Burger, head of Microsoft Research, teams up with experts to explore the challenges and promises of artificial intelligence. This first episode features Nicolò Fusi, also from Microsoft Research, and Subutai Ahmad from Numenta, discussing the true nature of machine intelligence.

Technical advancements are progressing at such a rapid pace that it can be difficult to define the future we are working towards. In this series, experts from various disciplines address the most pressing AI-related questions faced by technologists, policymakers, business leaders, and other stakeholders today. The goal is to amplify the shared understanding necessary to build a future where the transition to AI is a net benefit.

Comparing AI and the Human Brain

The speakers examine the fundamental differences between transformer-based language models and the human brain, which learns continuously. Subutai Ahmad emphasizes that the human brain constantly models its environment, allowing for rapid adaptation to changes through neurotransmitters. This capacity for continuous learning contrasts with current AI systems, which cannot update their knowledge as fluidly.

The discussion questions what intelligence truly means, where current models excel or fail, and what future AI systems might require to bridge the gap. Doug Burger and Subutai Ahmad have exchanged emails and discussed collaborations over the years, enriching their mutual understanding of the challenges posed by artificial intelligence.

A notable example is drawn from Jeff Hawkins' book On Intelligence, illustrating how the human brain learns continuously. This example highlights the ability of biological organisms to constantly adapt to their environment, a process that remains difficult to replicate in current AI systems.

Contributions from the Experts

Doug Burger

Doug Burger hosts the podcast seeking to understand how advancements in AI can transform our world. He is particularly interested in how these technologies can be both beneficial and problematic.

Nicolò Fusi

Nicolò Fusi, with his expertise in Bayesian methods and computational biology, brings a unique perspective on generative modeling and information theory applied to language models. He began by studying Gaussian processes and has moved away from kernel methods and Bayesian non-parametrics to focus more on language models.

Subutai Ahmad

Subutai Ahmad, with his background in computer science and cognitive psychology, focuses on applying principles of brain function to AI. He discusses the importance of neurotransmitters in human learning and the challenges this poses for current AI systems. Ahmad has been very interested in Jeff Hawkins' idea of deeply understanding how the brain works and finding ways to apply that to AI.

Towards a Smarter AI

The discussion continues with an analysis of transformers, the models currently dominating the AI landscape. The experts debate what these systems still need to achieve to come closer to human intelligence, particularly in terms of efficiency and sensorimotor representation. Large language models, while impressive, still fail to match the complexity and flexibility of human learning. Researchers agree that to bridge this gap, future AI systems will need to incorporate learning mechanisms more akin to those observed in the human brain.