Customer Reviews

The Mind's Arrows: Bayes Nets and Graphical Causal Models in Psychology

5 star:		(0)
4 star:		(1)
3 star:		(0)
2 star:		(0)
1 star:		(0)

 

 

This book could be considered to be a non-mathematical introduction to Bayesian networks, as applied to the field of cognitive psychology. It is readily apparent that the author does not believe that traditional statistical methods, such as factor and regression analysis, are unsuitable tools for the study of how the human mind gains an understanding and knowledge of the causal structure of the world. Bayesian networks are thus brought in to replace these methods.

The author begins the book by distinguishing between `graphical causal structures,' which include structures with feedback, from `causal Bayes nets', which do not. He concentrates on the latter in the book, believing that they have great utility in psychological theory, namely in studies of adult judgment, psychometrics, cognitive neuropsychology, and in developmental and social psychology. In particular, causal Bayes nets can be used to successfully model the 'theory theory' which was developed to understand how cognition develops in infants and children. Children are of course able to learn the causal structure of the world, and this ability can be captured by the use of causal Bayes nets. In the context of developmental psychology, the emphasis in the book is on how children obtain the ability to predict and control their environment, and not on how they obtain the ability to generate explicit causal explanations.

The author gives a brief history of cognitive psychology in order to motivate the ideas in the book. The discussion is intriguing, for it sheds light on the attitudes, biases, and the frequent extreme pessimism of practitioners in the field. Some cognitive psychologists for example held to the belief that humans are unable to conform to moral or rational standards, and that even if they are able to do this in some contexts or circumstances, a change in these circumstances will suppress this conformity. Science, the ultimate in rational endeavor, is in this view an "unstable oddity" that can only be sustained if sophisticated structures of social interaction are constructed.

The author though has a realistic assessment of practice, and points to the "child scientist" as being one that has a genuine desire for the understanding and control of the environment, is not worried about competition for jobs or tenure, and thus is not unduly biased by the need for them. The "child scientist" is to be distinguished from the "adult scientist," the latter of which is distracted by societal issues that force them out of their rational equilibrium.

Most interesting, because the author is a professional philosopher, is his statement that twentieth-century philosophy does not seem to permit any concept of the logic of discovery. Scientific inquiry or investigation, hypothesis testing, etc, cannot be described algorithmically. The author though describes these views as being "quaint," and offers as proof the progress that has been made in machine learning. The study of machines or 'computational systems' that can gain knowledge of the world via its sensory inputs and 'primitive abilities' is what the author has designated as 'android epistemology.' The author characterizes 'android epistemology' as being the most ambitious project of all in artificial intelligence. What he does not mention though is that major progress has been made in this project in the last decade. Along these same lines, a newcomer to artificial intelligence will hear mention of the 'frame problem,' which involves the need for the specification of not only what changes but also what does not change under a particular action. The frame problem has been the subject of considerable debate in the artificial intelligence community, and is also an issue in the developmental psychology of infants and children.

The construction of a causal Bayes net uses the causal Markov assumption and the notion of an acyclic directed graph. The author views causes, effects, etc. in terms of concepts: The concept of the causes of a feature or collection of features; the concept of the effects of a feature or collection of features, and most importantly causes are to be distinguished from covariates. To reduce computational complexity, edges can be reduced or connections with low probability can be eliminated. Variables with a common effect can be collapsed, as well as causes with distinct effects. Features that are mutually exclusive can be represented by an abstract variable. Variables that are intermediate between other variables can be deleted if adjustments are made, and variables can be refined and coarsened if needed. Prior knowledge can be used or omitted if desired.

A sizable portion of the book is devoted to the discussion of the efficacy of human judgment in causation. The author discusses experiments that test this efficacy, and how the data is interpreted with respect to the Rescorla-Wagner model. This model has dominated psychological theories of human and animal learning for many years, but the author discusses an example that indicates problems with this model. This is followed by a discussion of the Cheng model of human judgment of generative causal power. The Cheng model can be expressed as a Bayes net, and the author gives detailed discussion on just how effective this representation can be.

The most interesting chapter of the book concerns the use of neural networks to study the effects of brain lesions. A baseline neural network is constructed that is supposed to emulate the normal capacity of the brain. This network is then altered in order to model the functioning of a damaged brain. This strategy is particularly interesting, and very important considering the enormous efforts that are currently being made to connect experimental data to neural network architectures. The author quotes theorems that indicate that the transmission functions of the neural network are really arbitrary as far as the independence properties of the network are concerned. The "lesioning" of the neural network cannot eliminate any of these independencies. These results, which seem to cast doubt on the strategy of using lesioning, do not negate the use of Bayesian neural networks to study brain lesions. These theorems do not invalidate the use of Bayesian neural networks.