Carnegie Mellon University

Lectures & Colloquia

Spring 2020

Tuesday, January 21Philosophy Colloquium
Julia Haas, Rhodes College and Australian National University
Talk Title: Passing the Moral Turing Test
4:30 pm-6 pm, Baker Hall A51 (Giant Eagle Auditorium)

Abstract: The translation problem in moral AI asks how insights into human norms and values can be translated into a form suitable for implementation in artificial systems. I argue that if my answer to a question about the human mind is right, then the translation problem is more tractable than previously thought. Specifically, I argue that we can use principles from reinforcement learning to study human moral cognition, and that we can use principles from the resulting evaluative moral psychology to design artificial systems capable of passing the Moral Turing Test (Allen, 2000).

Thursday, January 23Philosophy Colloquium
Jorge Morales, Johns Hopkins University
Talk Title:  Does the World Look Flat? Sustained Representation of Perspectival Shape
4:30 pm-6 pm, Baker Hall A51 (Giant Eagle Auditorium)

Abstract: Arguably the most foundational principle in perception research is that our visual experience of the world goes beyond the retinal image. We effortlessly perceive the distal environment rather than the proximal stimulation it causes. Shape, in particular, may be the paradigm case of such “unconscious inference”: When a circular coin is rotated in depth, for example, we see the circular object it truly is, not the perspectival ellipse it projects on our eyes. Despite the ubiquity of shape constancy and the unusually broad set of contemporary philosophical debates that have a stake in this discussion, past and contemporary philosophers have struggled to describe what our experience of an object’s shape is like when its objective shape does not match its perspectival shape. Does a tilted coin look like a circular object, or like an elliptical object? Moreover, once our visual system infers that an elliptical projection arose from a distally circular object, do our minds still represent the rotated coin’s ellipticity? Is the retinal image discarded, as is assumed by most scientific models? To address these questions, I resort to the methods of visual science. If the elliptical perspectival shapes persist in the mind, objectively circular objects should, when rotated, impair search for objectively elliptical objects. Results from seven experiments demonstrate that this is so, indicating that perspectival shape representations indeed persist in our minds far longer than is traditionally assumed both in philosophy and cognitive science. Importantly, this case study on perspectival shape offers a glimpse into the cognitive architecture of the visual system as well as into the computations it relies on to represent the world around us. I end by reflecting on the nature of this cognitive architecture and on the computational efficiency brought about by persistent perspectival representations.

Monday, January 27Philosophy Colloquium
Sydney Levine, Massachusetts Institute of Technology
Talk Title: The structure of the moral mind
4:30 pm-6 pm, Baker Hall A53 (Steinberg Auditorium)

Abstract: In this talk, I demonstrate how my research on moral judgment puts normative and empirical concerns in dynamic exchange with one another.  Normative ethics can provide hypotheses for how to build and test empirical, computationally-grounded models of the moral mind.  These models, and their implications for AI engineering, then raise a new set of normative questions about how to develop and regulate AI.

First, I will show how ideas from moral philosophy inspire my empirical work.  I will discuss my work on the nature, structure, and developmental origin of moral rules through a case study in preschoolers' use of the "means principle".  I will then discuss a project that investigates a mechanism of moral decision-making ("universalization") that is used in situations where no moral rule exists.  Universalization (a version of Kant’s Categorical Imperative) asks us to consider what would happen if everyone behaved a certain way. This project integrates behavioral and developmental approaches to build a broadly computational account of universalization that makes fine-grained quantitative predictions.  Finally, I will discuss recent projects which draw out the normative implications of my empirical work for AI policy and regulation.

Monday, February 3Philosophy Colloquium
Paula Rubio-Fernandez, University of Oslo / Massachusetts Institute of Technology
Talk Title: Is there an egocentric bias in communication?  Experimental studies on Theory of Mind use
4:30 pm-6 pm, Baker Hall A53 (Steinberg Auditorium)

Abstract: Mastering communication takes more than mastering a language. Often, the point of a message is only implied by what we say. This means that, as speakers, we trust our listeners to read between the lines, and as listeners, we are willing to go beyond the literal to infer what the speaker intended to convey (Grice, 1975). Theoretical work on the nature of communication has long argued that these ‘mindreading’ inferences must rely on Theory of Mind: our ability to reason about other people’s mental states, such as their beliefs and intentions (Sperber & Wilson, 1986; Levinson, 2006; Tomasello, 2008). Even though language and Theory of Mind come together in communication, their relationship has been intensely contested, both in adult communication and in child development (for discussion, see Moore, 2017; Rubio-Fernandez, 2019; Rubio-Fernandez et al., 2019). In this talk, I will present the results of a series of studies investigating how Theory of Mind may be employed in referential communication. First, I will look at the issue of informativity, and whether speakers who are redundant (potentially violating the Gricean Maxim of Quantity) are failing to take their interlocutor’s perspective (a behavior that has been described as ‘egocentric’; Keysar, 2007). I will challenge this view and show how previous accounts have failed to appreciate the discriminatory value of redundant referential expressions (e.g., ‘Give me the blue star’ in a display in which the star is the only blue shape), focusing only on their informational value (i.e. whether there are other stars in the display that could justify the use of a color adjective). In the second part of my talk, I will look at egocentricity from the listener’s perspective and challenge the results of a standard task that has been taken to show the limits of Theory of Mind use in communication (the so-called ‘Director task’; Keysar et al., 2003). I will show, using both interactive and online versions of this task, how adults make rather sophisticated use of their Theory of Mind in referential communication, deriving pragmatic inferences not only to interpret a message, but also to update their common ground with the speaker. I will conclude by sketching a new line of research with which I plan to investigate the role of pragmatics in Theory of Mind development.


  • Grice, H. P. (1975). Logic and conversation. In P. Cole and J. L. Morgan (Eds.), Syntax and semantics, Vol. 3, Speech Acts (pp. 41–58). New York: Academic Press.
  • Keysar, B. (2007). Communication and miscommunication: The role of egocentric processes. Intercultural Pragmatics, 4, 71–84.
  • Keysar, B., Lin, S., & Barr, D. J. (2003). Limits on Theory of Mind use in adults. Cognition, 89, 25-41.
  • Levinson, S. C. (2006). Cognition at the heart of human interaction. Discourse Studies, 8, 85-93.
  • Moore, R. (2017). Gricean communication, language development, and animal minds. Philosophy Compass, 13, e12550.
  • Rubio-Fernandez, P. (2019). Theory of Mind. In C. Cummins and N. Katsos (Eds.). The Handbook of Experimental Semantics and Pragmatics (pp. 524–536). Oxford: Oxford University Press.
  • Rubio-Fernandez, P., Mollica, F., Ali, M. O., & Gibson, E. (2019). How do you know that? Automatic belief inferences in passing conversation. Cognition, 193, 104011.
  • Sperber, D. & Wilson, D. (1986/1995) Relevance: Communication and cognition. Oxford: Basil Blackwell.
  • Tomasello, M. (2008). Origins of Human Communication. Cambridge, MA: MIT Press.

Tuesday, February 18 - Philosophy Colloquium 
Francesca Zaffora Blando, Stanford University
Talk Title: Algorithmic randomness and Bayesian learning for computable agents 
4:30 pm-6 pm, Baker Hall A53 (Steinberg)

Abstract: Algorithmic randomness is a branch of computability theory that provides an analysis of the notion of a sequence displaying no effective patterns. In this talk, I will explore the effects of algorithmic randomness, taken to be a property of data streams, on the learning performance of computable Bayesian agents: i.e., agents whose priors are computable. 

First, I will show that several canonical algorithmic randomness notions can be characterised via suitably effectivised versions of Lévy’s Upward Theorem, a fundamental Bayesian convergence-to-the-truth result. Lévy’s theorem establishes that belief, in the form of a Bayesian agent’s best estimate of the true value of a random variable, aligns with the truth in the limit with probability one. The results I will present are a sharpening of Lévy’s classical result: I will show that, for computable Bayesian learners trying to approximate effective random variables, the truth-conducive data streams coincide with the algorithmically random ones. Thus, the algorithmically random data streams are exactly the ones along which a certain kind of inductive success is attainable. I will also consider the Blackwell-Dubins Theorem—a seminal Bayesian merging-of-opinions result—from this perspective, and discuss some preliminary findings. 

Second, I will discuss the use of algorithmic randomness to define notions of compatibility between the beliefs of computable Bayesian agents. I will explain how the resulting compatibility concepts relate to each other and to standard notions of agreement between measures (first and foremost, absolute continuity and effective variants thereof). Then, we will see that, for many canonical notions of algorithmic randomness, the notions of compatibility that they respectively induce lead to merging of opinions in the strong sense of the Blackwell-Dubins Theorem. 

These results suggest a fruitful interaction between the theory of algorithmic randomness and the study of inductive learning for computationally limited agents.

Friday, February 21 - Philosophy Colloquium
Regina Rini, York University
Talk Title: Moral Disagreement is Special
3:30 pm-4:45 pm, Baker Hall A53 (Steinberg)

Abstract: Most of the literature on moral disagreement is framed in strictly epistemic terms. I argue that this framing is misleading, as moral disagreement is unlike peer disagreement in other epistemic domains, owing to the special character of the moral domain. I defend the claim that disagreement with peers gives us reason to reduce confidence in disputed moral beliefs, but not for epistemic reasons. Rather, we have moral reason to do so. Reducing confidence in this way is morally required by recognition respect for the moral agency of the peer with whom we disagree.

Monday, February 24Philosophy Colloquium 
Ulrik Buchholtz Technische Universität Darmstadt
Talk Title: What rests on what, homotopically? 
4:30 pm-6 pm, Baker Hall A53 (Steinberg)

Abstract: Homotopy theory is becoming increasingly ubiquitous in many areas of pure mathematics. Part of the reason is the growing realization that the notion of equality of general mathematical objects needs to be enriched from a mere truth-value to a structured mathematical object in its own right. Homotopy type theory offers a language that formalizes this idea.

When Feferman asked in 1992, “What rests on what?”, this homotopical shift was not yet on the logician's radar. I will talk about the role homotopy type theory can play in the realms of proof-theoretic reductions and the foundations of mathematics. In the same year, Feferman also argued that “A little bit goes a long way,” in that logically weak systems can secure all scientifically applicable mathematics. Of course, the latter is a moving target: Soon, homotopy theory and higher groups may be essential in a physical theory (whether through string theory or something else). I will argue that by developing mathematics in homotopy type theory, we strengthen the case that a little bit indeed goes a long way.

Monday, March 2Philosophy Colloquium
Aybüke Özgün, ILLC, University of Amsterdam & Arch´e, University of St. Andrews
Talk Title: Refining Epistemic Logic
4:30 pm-6 pm, Baker Hall A53 (Steinberg)

Abstract: Epistemic logic is an umbrella term for a variety of modal logics whose main objects of study are knowledge and belief. As a field of study, epistemic logic uses mathematical tools to formalize, clarify, and address the questions that drive (formal) epistemology, and its applications extend not only to philosophy, but also to theoretical computer science, artificial intelligence, and economics. Research in epistemic logic has widely advanced based on the formal ground of normal modal logics and standard possible worlds semantics on relational structures as they provide a relatively easy way of modeling knowledge and belief. However, this mainstream approach is subject to well-known conceptual objections and open to extensions to better handle information. In this talk, I will focus on three features of the standard (relational) possible worlds semantics that call for refinement/enrichment and explore three directions, based on mathematically-richer settings, targeting the particular logical or philosophical issues in question.

The first direction is concerned with the relationship between notions of evidence, knowledge, and belief. The orthodox treatment of epistemic logics based on the standard relational semantics is not rich enough to talk about the evidential nature of acquired knowledge and belief. I argue that topological spaces emerge naturally as information structures if one not only seeks an easy way of modeling knowledge and belief, but also aims at representing evidence and its relationship to knowledge and belief. The topological approach enables fine-grained and refined representations of the aforementioned epistemic notions. The second target of analysis is the problem of logical omniscience. Standard possible worlds semantics models highly idealized reasoners, far from having realistic cognitive powers and bounds: the agents represented know/believe all logical truths, and know/believe all logical consequences of what they know/believe. The key idea I exploit to achieve non-omniscience focuses on topic- or subject matter-sensitivity of intentional mental states. Third, I will discuss logics of abstraction, meant to formalize the act of “abstracting away” the irrelevant features of an epistemic model. This phenomenon is pervasive in formal epistemology: when modeling epistemic scenarios, the modeler focuses on a set of relevant issues, and identifies situations that agree on all these issues, thus reducing the size and complexity of the model to manageable proportions. Our proposed logics internalize this process.

Finally, I will sketch how these themes fit in a unified research program and point to further extensions and applications in mathematical logic, philosophical logic, and formal learning theory.

Friday, March 20 - Center for Ethics and Policy Colloquium
Jonathan D. Moreno, University of Pennsylvania
Talk Title: Mind Wars: Brain Science and National Security
3:30-4:45pm, Baker Hall A53 (Steinberg)

Abstract:  My book Mind Wars examines the ethical dilemmas and fascinating history of brain science for military and intelligence applications. In this talk, I will discuss the longstanding interest and support of neuroscience by government and the roles of the national security community and university science departments in preparing the military and intelligence services for the changing nature of armed conflict.  I argue that the future of national security will depend partly on the practical applications of neuroscience to autonomous weapons, brain-computer interface, warfighter psychology, personal enhancement, intelligence analysis and many other technologies and modes of conflict.

Friday, March 27Philosophy Colloquium
Jean Baccelli, University of Pittsburgh
3:30-4:45pm, BH A53

Friday, April 24Philosophy Colloquium
Adina Roskies, Dartmouth College
3:30-4:45pm, BH A53

Fall 2019

Friday, September 13 - Center for Ethics and Policy Colloquium
Jennifer Prah Ruger, University of Pennsylvania
Talk Title: Global Health Justice and Governance
3:30pm-4:45pm, BH A53

Abstract: In a world beset by serious and unconscionable health disparities, by dangerous contagions that can circle our globalized planet in hours, and by a bewildering confusion of health actors and systems, humankind needs a new vision, a new architecture, new coordination among renewed systems to ensure central health capabilities for all. Professor Jennifer Prah Ruger's book Global Health Justice and Governance (Oxford University Press, 2018) lays out the critical problems facing the world today and offers a new theory of justice and governance as a way to resolve these seemingly intractable issues. In this seminar, Professor Prah Ruger will address the fundamental responsibility of society to ensure human flourishing; the central role that health plays in flourishing and how that places a unique claim on our public institutions and resources to ensure central health capabilities to reduce premature death and avoid preventable morbidities; and the new global health architecture that is needed in order to address staggering inequalities, imperiling epidemics, and inadequate systems.

Friday, September 20 - Philosophy Colloquium
Chris Barker, New York University
Talk Title: Natural language imperatives as actions
3:30pm-4:45pm, BH A53

Abstract: Imperatives ("Sit down!") are instructions for changing the world in a certain way. Various logics associated with Pratt, Hoare, Segerberg, and others formalize the behavior of computer programming languages in terms of actions, modeled as relations over states. I will argue that Segerberg's Dynamic Logic, in particular, is a promising candidate for describing natural language imperatives. For instance, "Sit down!" would denote the set of pairs such that b is a continuation of a in which the addressee of the imperative has sat down. Then conjoining imperatives corresponds to dynamic sequencing, so that "Sit down and shut up!" is the set of pairs such that there is some b within the denotation of "Sit down!" and in the denotation of "Shut up!". Disjunctions of imperatives naturally correspond to set union, which means that Dynamic Logic validates free choice inferences ("Eat an apple or eat a pear" can be obeyed by eating an apple), without validating Ross' Paradox (the set of actions denoted by "Mail this letter" will not include the set of actions denoted by "Mail this letter or burn it"). The right way to handle negation ("Don't be sad!") is less clear, but the treatment of negation in Krifka's more general conception of speech acts as index space shifters points the way.

Friday, November 8 - Philosophy Colloquium
Silvia De Toffoli, Princeton University
Talk Title: Diagrammatic Notations in Mathematical Proofs
3:30pm-4:45pm, BH A53

Abstract: The aim of my talk is to investigate the roles played by diagrams -- and in particular of diagrammatic notations -- in the context of proofs in mathematics.  I begin by considering the problem of how to characterize mathematical diagrams in the first place and propose a technical definition to delineate the phenomenon under investigation.  The focus of the talk is on the features of diagrams that underwrite their ability to play a role in the inferential structure of certain proofs. My main thesis is that in some cases not only proof presentations but the proofs themselves are dependent (in a way I will specify) on certain features of the notations they deploy -- I call these constitutive features.  Throughout the talk, I use examples of diagrams taken from different areas of contemporary mathematics.

Friday, November 15 - Alumni Colloquium
Conor Mayo-Wilson, University of Washington
Talk Title: Qualitative Objective Likelihoodism
3:30pm-4:45pm, BH A53

Abstract: In debates about the foundations of statistics, likelihoodism is roughly the thesis that all information about an unknown statistical parameter is contained in the likelihood function. Although likelihoodists often advertise their methods as more "objective" than Bayesian ones, the nature of that "objectivity" is often not entirely clear and its importance even less so. We first show that core likelihoodist theses can be understood as characterizing when all Bayesians will agree in a different sense. We then argue that this agreement is a better explication of the sense in which likelihoodist methods are objective. Finally, we show that the same results extend to qualitative probabilistic settings, in which prior and posterior "probabilities" are represented by qualitative orderings over the hypothesis and outcome spaces.