Unifying the Mind: Cognitive Representations as Graphical Models

by David Danks

A novel proposal that the unified nature of our cognition can be partially explained by a cognitive architecture based on graphical models. Our ordinary, everyday thinking requires an astonishing range of cognitive activities, yet our cognition seems to take place seamlessly. We move between cognitive processes with ease, and different types of cognition seem to share information readily. In this book, David Danks proposes a novel cognitive architecture that can partially explain two aspects of human its relatively integrated nature and our effortless ability to focus on the relevant factors in any particular situation. Danks argues that both of these features of cognition are naturally explained if many of our cognitive representations are understood to be structured like graphical models. The computational framework of graphical models is widely used in machine learning, but Danks is the first to offer a book-length account of its use to analyze multiple areas of cognition. Danks demonstrates the usefulness of this approach by reinterpreting a variety of cognitive theories in terms of graphical models. He shows how we can understand much of our cognition—in particular causal learning, cognition involving concepts, and decision making—through the lens of graphical models, thus clarifying a range of data from experiments and introspection. Moreover, Danks demonstrates the important role that cognitive representations play in a unified understanding of cognition, arguing that much of our cognition can be explained in terms of different cognitive processes operating on a shared collection of cognitive representations. Danks's account is mathematically accessible, focusing on the qualitative aspects of graphical models and separating the formal mathematical details in the text.

Genres: Philosophy

304 pages, Hardcover

First published September 5, 2014

Reviews

[KC · Rating 4 out of 5]

The central thesis of this book is that the mind makes use of (causal) graphical models as a shared underlying representation for many important cognitive processes, including causal reasoning, decision making, and categorization (a chapter is devoted to each of these). This idea predicts, for example, that learning causal features of some concept while attempting to categorize it will allow you to later figure out how to intervene on this learned causal structure to produce a desired outcome. I think the case for using graphical models for causal reasoning and decision-making has been well-established previously, but the idea of using them for categorization was new to me. It's interesting that several theories of concept learning can be cast as graphical models, but I didn't think that a very successful case was made for why this is a useful interpretation. For example, it seems equally plausible to me that cateogrized concepts are represented using a different structure, and once a concept is learned it can be incorporated into a graphical model that would allow for causal inference. However, the overall idea of a shared representation store that different cognitive processes manipulate in different ways is interesting and I'm glad I took the time to learn about this perspective.

My favorite aspect of this book was Danks' commitment to clearly contextualizing his ideas. He devotes significant amounts of space to describing the relationships between various cognitive theories and attempting to clarify which competing hypotheses are necessarily in conflict with each other and which are perhaps compatible. He is also refreshingly honest about the limitations and potential weaknesses of his own central thesis. I think the framework Danks presents for thinking about this field will stick with me, even if some of the new ideas he proposes do not.

[Sven · Rating 4 out of 5]

Fun and refreshing. Especially in the beginning there is a great part on the philosophy behind the approach and on explanation in cognitive science. The author makes explicit what this framework can add to current cognitive science, which is nice to read. The actual content of the work that has been done though was less engaging for me.