Algorithmic approaches to ecological rationality in humans and machines

Abstract

In a complex and ever-changing world, how do humans reason as intelligently as they do—especially given limited energy, data, and time? How can understanding this guide us toward building better artificially intelligent systems? Bayesian models provide a normative account of rational behavior. Although computing rational responses via exact Bayesian inference is expensive, empirical findings show that human behavior is often consistent with these rational responses. This seems to indicate that an efficient and accurate inference engine underlies human cognition. However, in several notable cases, humans display ‘cognitive biases’, where their judgments deviate systematically from exact Bayesian inference. How can we reconcile these contradicting findings? This thesis provides a reconciliation by building on the insight that humans are not general purpose computers: we are instead ‘ecologically rational’, adapting to structure in our environments to make the best use of limited computational resources. I first discuss algorithms for approximating exact Bayesian inference within limited computational resources. These reduce the costs of inference by leveraging underlying environmental structure through ‘amortization’: the adaptive re-use of previous computations. However, amortization can lead to errors when the current query is not representative of past experience. I demonstrate that these errors replicate several human cognitive biases, and test new predictions with behavioral experiments. Finally, I show that amortization also gives rise to ecologically rational behaviors in machine learning, and demonstrate how this can be leveraged to artificially engineer new kinds of intelligent behaviors, like causal reasoning and compositional language representation. This also provides new insights into how these central tenets of intelligence manifest in humans. By taking an algorithmic approach to ecological rationality—that is, by making explicit claims about how it can be implemented at the level of computational processes—this thesis develops new models for human probabilistic inference that can explain both its remarkable successes as well as its seeming failures, and also suggests new avenues toward machines with human-like intelligence.