Our Work
This paper is about how an agent should rationally update her probabilistic beliefs when her conceptual space (modelled as an algebra of propositions) grows. This is not like typical cases of learning, which are cases in which an agent comes to revise her beliefs for propositions about which she was already aware. We investigate whether the learning rules for the typical cases of learning can be extended to the case of conceptual growth.
Read MoreHere HMI CI Katie Steele argues that on a certain way of modelling an agent's preferences and understanding her "time preferences", exponential time discounting is uniquely rational. However, if "time preferences" are understood differently, then exponential time discounting is not uniquely rational. This helps in understanding why the prescription of exponential time discounting has many defenders but also many detractors.
Read MoreIn semester 2, HMI RFs Alban Grastien and Atoosa Kasirzadeh are teaching a course entitled "Advanced Topics on Artificial Intelligence" in the Research School of Computer Science at the ANU. This course presents some of the techniques developed in AI for decision making under uncertainty, and introduces the variety of moral and sociological implications of these decisions.
Read MoreThis paper identifies a new role for mathematics in scientific practice. Atoosa calls this the "bridging'' role of mathematics, according to which mathematics acts as a connecting scheme in our explanatory reasoning about why and how two different descriptions of an empirical phenomenon relate to each other.
Read MoreSeth Lazar joined the Templeton World Charity Foundation Diverse Intelligences Summer Institute as faculty, giving talks on the moral and political epistemology of data and AI, and the Value of Explanations.
Read MoreThis talk was given at a conference on Holly Smith’s book, Making Morality Work, held at Rutgers on October 18, 2019. I argued that Making Morality Work poses the problem that moral theories must be 'usable', but then offers a solution that only partly solves it. I offered a way to extend the solution, but argued that even that only partly solves the problem, and that we can’t stop there.
Read MoreWe put forth an analysis of actual causation. The analysis centers on the notion of a causal model that provides only partial information as to which events occur, but complete information about the dependences between the events. The basic idea is this: c causes e just in case there is a causal model that is uninformative on e and in which e will occur if c does. Notably, our analysis has no need to consider what would happen if c were absent. We show that our analysis captures more causal scenarios than any counterfactual account to date.
Read MoreWhat does a rational agent or an AI learn from a conditional? Günther (2018) proposed a method for the learning of indicative conditionals. Here, we amend the method by a distinction between indicative and subjunctive conditionals. As a result, the method covers the learning of subjunctive conditionals as well.
Read MoreThis talk was given to the effective altruism society at ANU on October 1, 2019. I described ethical problems associated with the project of designing ethical self-driving cars, what makes the project especially difficult, what we might do about it, and why those concerned with doing the most good should care.
Read MoreSeth Lazar and Colin Klein question the value of basing design decisions for autonomous vehicles on massive online gamified surveys. Sometimes the size of big data can't make up for what it omits.
Read MoreThe paper extends a deep learning model known as graph neural networks and uses it to learn generalised heuristics. We show that these heuristics generalise to problems with different goals, larger problems, and even problems from different domains than those we trained on. This is the first paper that successfully learns domain-independent heuristics.
Read MoreThis paper considers strategies for making decisions in the face of severe uncertainty, when one's beliefs are best represented by a set of probability functions over the possible states of the world (as opposed to a single precise probability function). The question is whether one can employ a decision strategy that does not have the disadvantage of making one vulnerable to sure loss in sequential-decision scenarios.
Read More
