Towards Machines that Trust: AI Agents Learn to Trust in the Trust Game

Widely considered a cornerstone of human morality, trust shapes many aspects of human social interactions. In this work, we present a theoretical analysis of the $\textit{trust game}$, the canonical task for studying trust in behavioral and brain sciences, along with simulation results supporting our analysis. Specifically, leveraging reinforcement learning (RL) to train our AI agents, we systematically investigate learning trust under various parameterizations of this task. Our theoretical analysis, corroborated by the simulations results presented, provides a mathematical basis for the emergence of trust in the trust game.

Cognitive Models as Simulators: The Case of Moral Decision-Making

To achieve desirable performance, current AI systems often require huge amounts of training data. This is especially problematic in domains where collecting data is both expensive and time-consuming, e.g., where AI systems require having numerous interactions with humans, collecting feedback from them. In this work, we substantiate the idea of $\textit{cognitive models as simulators}$, which is to have AI systems interact with, and collect feedback from, cognitive models instead of humans, thereby making their training process both less costly and faster. Here, we leverage this idea in the context of moral decision-making, by having reinforcement learning (RL) agents learn about fairness through interacting with a cognitive model of the Ultimatum Game (UG), a canonical task in behavioral and brain sciences for studying fairness. Interestingly, these RL agents learn to rationally adapt their behavior depending on the emotional state of their simulated UG responder. Our work suggests that using cognitive models as simulators of humans is an effective approach for training AI systems, presenting an important way for computational cognitive science to make contributions to AI.

Bringing Order to the Cognitive Fallacy Zoo

In the eyes of a rationalist like Descartes or Spinoza, human reasoning is flawless, marching toward uncovering ultimate truth. A few centuries later, however, culminating in the work of Kahneman and Tversky, human reasoning was portrayed as anything but flawless, filled with numerous misjudgments, biases, and cognitive fallacies. With further investigations, new cognitive fallacies continually emerged, leading to a state of affairs which can fairly be characterized as the cognitive fallacy zoo! In this largely methodological work, we formally present a principled way to bring order to this zoo. We introduce the idea of establishing implication relationships (IRs) between cognitive fallacies, formally characterizing how one fallacy implies another. IR is analogous to, and partly inspired by, the fundamental concept of reduction in computational complexity theory. We present several examples of IRs involving experimentally well-documented cognitive fallacies: base-rate neglect, availability bias, conjunction fallacy, decoy effect, framing effect, and Allais paradox. We conclude by discussing how our work: (i) allows for identifying those pivotal cognitive fallacies whose investigation would be the most rewarding research agenda, and importantly (ii) permits a systematized, guided research program on cognitive fallacies, motivating influential theoretical as well as experimental avenues of future research.

Over-representation of Extreme Events in Decision-Making: A Rational Metacognitive Account

The Availability bias, manifested in the over-representation of extreme eventualities in decision-making, is a well-known cognitive bias, and is generally taken as evidence of human irrationality. In this work, we present the first rational, metacognitive account of the Availability bias, formally articulated at Marr's algorithmic level of analysis. Concretely, we present a normative, metacognitive model of how a cognitive system should over-represent extreme eventualities, depending on the amount of time available at its disposal for decision-making. Our model also accounts for two well-known framing effects in human decision-making under risk---the fourfold pattern of risk preferences in outcome probability (Tversky & Kahneman, 1992) and in outcome magnitude (Markovitz, 1952)---thereby providing the first metacognitively-rational basis for those effects. Empirical evidence, furthermore, confirms an important prediction of our model. Surprisingly, our model is unimaginably robust with respect to its focal parameter. We discuss the implications of our work for studies on human decision-making, and conclude by presenting a counterintuitive prediction of our model, which, if confirmed, would have intriguing implications for human decision-making under risk. To our knowledge, our model is the first metacognitive, resource-rational process model of cognitive biases in decision-making.

Converting Cascade-Correlation Neural Nets into Probabilistic Generative Models

Humans are not only adept in recognizing what class an input instance belongs to (i.e., classification task), but perhaps more remarkably, they can imagine (i.e., generate) plausible instances of a desired class with ease, when prompted. Inspired by this, we propose a framework which allows transforming Cascade-Correlation Neural Networks (CCNNs) into probabilistic generative models, thereby enabling CCNNs to generate samples from a category of interest. CCNNs are a well-known class of deterministic, discriminative NNs, which autonomously construct their topology, and have been successful in giving accounts for a variety of psychological phenomena. Our proposed framework is based on a Markov Chain Monte Carlo (MCMC) method, called the Metropolis-adjusted Langevin algorithm, which capitalizes on the gradient information of the target distribution to direct its explorations towards regions of high probability, thereby achieving good mixing properties. Through extensive simulations, we demonstrate the efficacy of our proposed framework.

Neural Implementation of Probabilistic Models of Cognition

Bayesian models of cognition hypothesize that human brains make sense of data by representing probability distributions and applying Bayes' rule to find the best explanation for available data. Understanding the neural mechanisms underlying probabilistic models remains important because Bayesian models provide a computational framework, rather than specifying mechanistic processes. Here, we propose a deterministic neural-network model which estimates and represents probability distributions from observable events --- a phenomenon related to the concept of probability matching. Our model learns to represent probabilities without receiving any representation of them from the external world, but rather by experiencing the occurrence patterns of individual events. Our neural implementation of probability matching is paired with a neural module applying Bayes' rule, forming a comprehensive neural scheme to simulate human Bayesian learning and inference. Our model also provides novel explanations of base-rate neglect, a notable deviation from Bayes.

Managing Uncertainty in Rule Based Cognitive Models

An experiment replicated and extended recent findings on psychologically realistic ways of modeling propagation of uncertainty in rule based reasoning. Within a single production rule, the antecedent evidence can be summarized by taking the maximum of disjunctively connected antecedents and the minimum of conjunctively connected antecedents. The maximum certainty factor attached to each of the rule's conclusions can be sealed down by multiplication with this summarized antecedent certainty. Heckerman's modified certainty factor technique can be used to combine certainties for common conclusions across production rules.