Cover Image for Theory in the Age of Unsafe AI
Cover Image for Theory in the Age of Unsafe AI
Avatar for AI Safety Nepal
Presented by
AI Safety Nepal
5 Going

Theory in the Age of Unsafe AI

Google Meet
Registration
Welcome! To join the event, please register below.
About Event

​The world is racing towards building powerful artificial intelligences that may pose major risks to the safety and well-being of humanity. This race is primarily engineering-driven, i.e., modern neural networks are grown through an increasingly complex alchemy of engineering decisions on architectural design, learning algorithms, dataset curation, and so on.

We lack a scientific and theoretical foundation to understand why this works and to decipher the mechanisms behind these AI systems, which is crucial for trying to understand, steer, and make these systems safe.

In this talk, Nischal will:
• Draw a historical parallel between the current situation and the invention of the steam engine and the subsequent development of the theory of thermodynamics that was required to render it safe.
• Present a bird's-eye survey of the current frontier in the theory of learning.
• Share his outlook for the future of theory and its application to safety.

​Speaker:

​Nischal Mainali is the Research Lead for Learning Theory at Principles of Intelligence. He works on theoretical models of learning and representations in modern neural networks with an eye towards interpretability. He approaches this with the exactly solvable tradition of learning dynamics theory in collaboration with Principia and using tools and methods from statistical physical to study feature learning in nonlinear models. He did his PhD in theoretical neuroscience, applying mathematical ideas to study spatial cognition in the mammalian brain.

Avatar for AI Safety Nepal
Presented by
AI Safety Nepal
5 Going