PhD Candidate, Department of Logic and Philosophy of Science, University of California, Irvine

Daniel’s Question: How should we create Artificial General Intelligence?

Keystone Title: An Introduction to Solomonoff Induction With an Application to Scientific Confirmation Theory

Abstract: Daniel pursues three main goals in this Keystons: to introduce readers to the Bayesian sequence prediction framework; to introduce Solomonoff’s prior (and provide a breif summary of arguments for why it is universally optimal); and to demonstrate Solomonoff’s prior solves the Zero Prior Problem (the problem that if one starts out with a belief of 0 in any hypothesis, then no matter how much evidence one observes in favor of the hypothesis, one will always be certain that the hypothesis is correct) in scientific confirmation theory. In doing so Herrmann aims to strengthen the claim that Solomonoff’s prior (a set of initial beliefs) is the universally optimal way to predict sequences.

Tell us about your Experiential Learning. How was your EL related to your Question or Keystone?
For my Experiential Learning, I did an internship at BenchSci, a bioscience start-up. It was directly related to my Question, How should we create artificial general intelligence?, which looked at both the societal/ethical and technical challenges of building an AGI. I developed both my intuition and technical knowledge about machine learning, which is one of the main contemporary approaches to AI.

How did you find your EL opportunity?
I was invited by someone I knew through Quest to attend the Rotman School of Management’s Machine Learning and the Market for Intelligence conference in Toronto. At the conference there was a startup booth session, where machine-learning startups displayed their products and business models. I spent time at almost every booth, and maintained relationships with some of the people I met. As summer came closer, I asked about internship opportunities, and secured a position with one of the startups. I used code I wrote in one of my Quest courses to demonstrate my capabilities.

How did your EL inform the path you’re on now?
For one thing, I realized I didn’t really want to work in a start-up! At the time, I was considering going into either technical philosophy or machine learning, and this experience pushed me towards the former. The intuition and know-how around neural networks and statistical prediction that I built during my EL have really helped me in my current PhD research, where I evaluate the practices and methodology of science, including neural networks.

Did you have any major revelations during your EL?
Yes! I came to the conclusion that machine learning is less advanced (powerful/developed) than I thought, but more advanced (powerful/developed) than most people think. That is, it is incredibly and surprisingly successful at what it does (function approximation), and yet is not a solution for all of the problems of AI. This was a major insight, and helped guide my interest towards understanding other approaches to AI and reasoning in general. I wouldn’t have been able to appreciate machine learning’s power or limitations without getting my hands dirty.