Interview, Fireside Chat
What is Statistics? (Michael I. Jordan) | AI Podcast Clips
- Decision-making is expected to require assumptions defining error and probabilities to minimize the likelihood of incorrect conclusions, with the probability of continued errors predicted to decrease over time when statistical principles are followed.
- The identification of underlying mechanisms from observed outcomes is anticipated to evolve into an "inverse problem," while the field analyzing census data for government policy is predicted to be formally named "statistics."
- Formalized statistics is expected to have developed alongside game theory and decision theory in the 1930s, with future curricula planned to teach decision theory as a precursor to Bayesian or frequentist methods.
- The "James Stein estimation" concept is predicted to present initial comprehension challenges requiring time to understand, while the conceptual relationship between Bayesian and frequentist approaches is expected to remain a persistent duality.
- Bayesian and frequentist approaches are expected to yield divergent answers in specific practices despite occasional similarities, with frequentist methods favored for software reliability guarantees of 95% correctness across unknown datasets and Bayesian methods prioritizing specific obtained data.
- A conflict between the need for frequentist guarantees and the Bayesian focus on specific data is expected to necessitate a blending of the two approaches, with empirical Bayes methods planned to plug estimates into the Bayesian framework in a mathematically assured manner.
- The Bayesian perspective is expected to enable the incorporation of human expertise into analysis under mathematical reassurance, contrasting with frequentist views that average over all possible data sets.
- The "false discovery rate" is expected to differ from standard metrics like precision or recall by measuring the fraction of bad discoveries among those announced, representing an inverse direction to classical frequentist thinking.
- Some quantities required for false discovery rate calculations are anticipated to be estimable in reasonable ways despite the need for priors, with the argument for this approach expected to have emerged from Robbins around 1960 and developed through the work of Brad Efron, Ben Yaminin, and John Storey.