Explaining Image Classifiers

We focus on explaining image classifiers, taking the work of Mothilal et al. [2021] (MMTS) as our point of departure. We observe that, although MMTS claim to be using the definition of explanation proposed by Halpern [2016], they do not quite do so. Roughly speaking, Halpern's definition has a necessity clause and a sufficiency clause. MMTS replace the necessity clause by a requirement that, as we show, implies it. Halpern's definition also allows agents to restrict the set of options considered. While these difference may seem minor, as we show, they can have a nontrivial impact on explanations. We also show that, essentially without change, Halpern's definition can handle two issues that have proved difficult for other approaches: explanations of absence (when, for example, an image classifier for tumors outputs "no tumor") and explanations of rare events (such as tumors).

Subjective Causality

We show that it is possible to understand and identify a decision maker's subjective causal judgements by observing her preferences over interventions. Following Pearl [2000], we represent causality using causal models (also called structural equations models), where the world is described by a collection of variables, related by equations. We show that if a preference relation over interventions satisfies certain axioms (related to standard axioms regarding counterfactuals), then we can define (i) a causal model, (ii) a probability capturing the decision-maker's uncertainty regarding the external factors in the world and (iii) a utility on outcomes such that each intervention is associated with an expected utility and such that intervention $A$ is preferred to $B$ iff the expected utility of $A$ is greater than that of $B$. In addition, we characterize when the causal model is unique. Thus, our results allow a modeler to test the hypothesis that a decision maker's preferences are consistent with some causal model and to identify causal judgements from observed behavior.

Inference for Probabilistic Dependency Graphs

Probabilistic dependency graphs (PDGs) are a flexible class of probabilistic graphical models, subsuming Bayesian Networks and Factor Graphs. They can also capture inconsistent beliefs, and provide a way of measuring the degree of this inconsistency. We present the first tractable inference algorithm for PDGs with discrete variables, making the asymptotic complexity of PDG inference similar that of the graphical models they generalize. The key components are: (1) the observation that, in many cases, the distribution a PDG specifies can be formulated as a convex optimization problem (with exponential cone constraints), (2) a construction that allows us to express these problems compactly for PDGs of boundeed treewidth, (3) contributions to the theory of PDGs that justify the construction, and (4) an appeal to interior point methods that can solve such problems in polynomial time. We verify the correctness and complexity of our approach, and provide an implementation of it. We then evaluate our implementation, and demonstrate that it outperforms baseline approaches. Our code is available at http://github.com/orichardson/pdg-infer-uai.

Joint Behavior and Common Belief

For over 25 years, common belief has been widely viewed as necessary for joint behavior. But this is not quite correct. We show by example that what can naturally be thought of as joint behavior can occur without common belief. We then present two variants of common belief that can lead to joint behavior, even without standard common belief ever being achieved, and show that one of them, action-stamped common belief, is in a sense necessary and sufficient for joint behavior. These observations are significant because, as is well known, common belief is quite difficult to achieve in practice, whereas these variants are more easily achievable.

Causal Models with Constraints

Causal models have proven extremely useful in offering formal representations of causal relationships between a set of variables. Yet in many situations, there are non-causal relationships among variables. For example, we may want variables $LDL$, $HDL$, and $TOT$ that represent the level of low-density lipoprotein cholesterol, the level of lipoprotein high-density lipoprotein cholesterol, and total cholesterol level, with the relation $LDL+HDL=TOT$. This cannot be done in standard causal models, because we can intervene simultaneously on all three variables. The goal of this paper is to extend standard causal models to allow for constraints on settings of variables. Although the extension is relatively straightforward, to make it useful we have to define a new intervention operation that $disconnects$ a variable from a causal equation. We give examples showing the usefulness of this extension, and provide a sound and complete axiomatization for causal models with constraints.

A Causal Analysis of Harm

As autonomous systems rapidly become ubiquitous, there is a growing need for a legal and regulatory framework to address when and how such a system harms someone. There have been several attempts within the philosophy literature to define harm, but none of them has proven capable of dealing with with the many examples that have been presented, leading some to suggest that the notion of harm should be abandoned and "replaced by more well-behaved notions". As harm is generally something that is caused, most of these definitions have involved causality at some level. Yet surprisingly, none of them makes use of causal models and the definitions of actual causality that they can express. In this paper we formally define a qualitative notion of harm that uses causal models and is based on a well-known definition of actual causality (Halpern, 2016). The key novelty of our definition is that it is based on contrastive causation and uses a default utility to which the utility of actual outcomes is compared. We show that our definition is able to handle the examples from the literature, and illustrate its importance for reasoning about situations involving autonomous systems.

Quantifying Harm

In a companion paper (Beckers et al. 2022), we defined a qualitative notion of harm: either harm is caused, or it is not. For practical applications, we often need to quantify harm; for example, we may want to choose the lest harmful of a set of possible interventions. We first present a quantitative definition of harm in a deterministic context involving a single individual, then we consider the issues involved in dealing with uncertainty regarding the context and going from a notion of harm for a single individual to a notion of "societal harm", which involves aggregating the harm to individuals. We show that the "obvious" way of doing this (just taking the expected harm for an individual and then summing the expected harm over all individuals can lead to counterintuitive or inappropriate answers, and discuss alternatives, drawing on work from the decision-theory literature.

From Outcome-Based to Language-Based Preferences

We review the literature on models that try to explain human behavior in social interactions described by normal-form games with monetary payoffs. We start by covering social and moral preferences. We then focus on the growing body of research showing that people react to the language in which actions are described, especially when it activates moral concerns. We conclude by arguing that behavioral economics is in the midst of a paradigm shift towards language-based preferences, which will require an exploration of new models and experimental setups.

The brain as a probabilistic transducer: an evolutionarily plausible network architecture for knowledge representation, computation, and behavior

We offer a general theoretical framework for brain and behavior that is evolutionarily and computationally plausible. The brain in our abstract model is a network of nodes and edges. Although it has some similarities to standard neural network models, as we show, there are some significant differences. Both nodes and edges in our network have weights and activation levels. They act as probabilistic transducers that use a set of relatively simple rules to determine how activation levels and weights are affected by input, generate output, and affect each other. We show that these simple rules enable a learning process that allows the network to represent increasingly complex knowledge, and simultaneously to act as a computing device that facilitates planning, decision-making, and the execution of behavior. By specifying the innate (genetic) components of the network, we show how evolution could endow the network with initial adaptive rules and goals that are then enriched through learning. We demonstrate how the developing structure of the network (which determines what the brain can do and how well) is critically affected by the co-evolved coordination between the mechanisms affecting the distribution of data input and those determining the learning parameters (used in the programs run by nodes and edges). Finally, we consider how the model accounts for various findings in the field of learning and decision making, how it can address some challenging problems in mind and behavior, such as those related to setting goals and self-control, and how it can help understand some cognitive disorders.

Reasoning About Causal Models With Infinitely Many Variables

Generalized structural equations models (GSEMs) [Peters and Halpern 2021], are, as the name suggests, a generalization of structural equations models (SEMs). They can deal with (among other things) infinitely many variables with infinite ranges, which is critical for capturing dynamical systems. We provide a sound and complete axiomatization of causal reasoning in GSEMs that is an extension of the sound and complete axiomatization provided by Halpern [2000] for SEMs. Considering GSEMs helps clarify what properties Halpern's axioms capture.