Abstract
Bayesian Models of Cognition: Reverse Engineering the Mind by Griffiths, Chater, and Tenenbaum is a comprehensive account of human cognition as probabilistic inference. The book promotes the view that many aspects of human thought, including perception, reasoning, learning, and decision-making, can be understood as approximate Bayesian inference, in which prior beliefs are updated in light of new evidence. While the idea of reverse-engineering cognition has a long tradition in cognitive science, dating back to David Marr's computational-level analysis and John R. Anderson's rational analysis, this volume is distinctive in formalizing these ideas within a general Bayesian framework and systematically applying them across multiple cognitive domains.
The book is organized into two major parts. The first introduces the foundations of Bayesian modeling, including probability theory, Bayes’ rule, graphical models, and inference techniques such as sampling and variational approximations. Chapter by chapter, the authors demonstrate how these formal tools are linked to cognitive phenomena and used to explain inductive reasoning, categorization, and learning under uncertainty. The second part shifts focus to more advanced topics, including hierarchical and nonparametric Bayesian models, probabilistic programming, intuitive theories (such as intuitive physics and intuitive psychology, i.e., theory of mind), and resource-rational cognition. By linking these topics to the foundations presented in the early chapters, the book maintains a coherent theoretical framework, avoiding the common problem of presenting cutting-edge methods as disconnected applications.
Rather than introducing reverse engineering as a new methodology, the book operationalizes it with precise probabilistic models. The authors start from the computational problems humans face, such as inferring causal structure or predicting outcomes, and derive optimal solutions using Bayesian principles. Human behavior is then evaluated relative to these idealized models, often revealing that cognition approximates rational inference under constraints. This approach situates the work within Marr's computational-level framework while also engaging with questions about algorithmic implementation and cognitive limitations.
A key strength of the book is its systematic attempt to apply a single formal framework across diverse domains of cognition. By framing cognition as probabilistic inference, the authors provide a common language for modeling phenomena ranging from concept learning and language acquisition to social reasoning. Although unifying cognition under a single framework is a longstanding ambition in cognitive science, this volume offers one of the most comprehensive and formally explicit realizations of that goal within the Bayesian paradigm. The result is both theoretically elegant and practically useful, allowing researchers to employ similar modeling techniques across distinct areas.
The book also integrates theory, methodology, and empirical application. The authors go beyond abstract exposition, demonstrating how to construct generative models and test them against behavioral data. Case studies illustrate how relatively simple probabilistic assumptions can yield rich and sometimes counterintuitive predictions about cognition. This balance of formal rigor and empirical relevance sets the book apart from more purely mathematical treatments of Bayesian methods. Additionally, the text engages with contemporary developments such as probabilistic programming and resource-rational analysis, showing how approximate inference methods, such as sampling, may correspond to cognitive processes. This strengthens the bridge between normative Bayesian models and psychological plausibility.
Despite these strengths, the book has some limitations. At a conceptual level, while the book engages seriously with some core challenges for Bayesian models, particularly the computational intractability of exact inference and the resulting need for approximation, it places less emphasis on deeper representational issues. The discussion of sampling methods and resource-rationality provides a plausible account of how idealized Bayesian inference may be implemented under cognitive constraints, and thus addresses an important gap between normative theory and psychological plausibility. However, the success of many models depends critically on the specification of structured hypothesis spaces and priors, which are often carefully designed rather than derived. The question of how such hypothesis spaces are themselves learned, discovered, or constrained, often referred to as the framing problem, remains only partially addressed. While this limitation reflects a broader open challenge for Bayesian approaches to cognition rather than a shortcoming unique to this volume, its relative lack of emphasis leaves an important aspect of the framework underexplored.
The book is also technically demanding: readers are expected to be comfortable with formal notation, abstract modeling, and computational thinking. It is well suited for self-study but requires substantial background knowledge and sustained effort. Those seeking practical guidance may find fewer step-by-step modeling instructions or programming exercises than in typical machine learning texts, and will likely need supplemental resources for implementation. Consequently, the book is best suited for advanced undergraduates, graduate students, or researchers with a strong quantitative background. Its modular structure allows instructors to select topics for courses in computational cognitive science or artificial intelligence, and its breadth makes it a valuable reference for interdisciplinary research.
In comparison to other works, Bayesian Models of Cognition occupies a distinctive niche. It differs from Bishop's Pattern Recognition and Machine Learning in that it prioritizes the interpretation and explanation of human thought rather than algorithmic performance and predictive accuracy. A different contrast can be drawn with Gelman et al.'s Bayesian Data Analysis, which is primarily concerned with statistical inference and data analysis rather than the development of cognitive theory. Relative to general cognitive science textbooks, this volume offers substantially greater formal precision and depth, albeit at the cost of accessibility. Finally, in contrast to edited handbooks, it presents a unified theoretical perspective rather than a collection of disparate viewpoints.
In conclusion, Griffiths, Chater, and Tenenbaum have produced a landmark volume that both synthesizes and advances Bayesian cognitive science. The book's central achievement lies in articulating a coherent vision of the mind as a probabilistic inference system, applying formal models consistently across domains, and connecting theory with empirical evidence. While technical complexity may limit accessibility for some readers, and some foundational challenges remain open, the depth, scope, and intellectual ambition make it an essential resource for anyone seeking to understand the computational foundations of human cognition.
