Everyone Interested in Cognitive Science Should Read This

In the late 1970s and early 1980s, it was clear that David Marr was one of cognitive science's brightest young minds. His philosophical, computational, and mathematical approaches to the problems presented to researchers by vision have been incredibly influential. Sadly, he was diagnosed with leukemia in the late 70s, and died in 1980 at the age of 35 (here is a short biography). As he was dying, he wrote a book that "redefined and revitalized the study of human and machine vision." The book, titled Vision: A Computational Investigation into the Human Representation and Processing of Visual Information, was published in 1982. The book is so good that you have to wonder what great things Marr might have done if he hadn't died so young. Much of Vision betrays the urgency with which he wrote it, as it often appears to be written in the form of edited personal notes, and much of it is difficult to read. I remember reading it for the first time and not understanding a word after about midway through Chapter 2. After a course on vision, I picked it back up, and it then made sense, but it still wasn't an easy read.

But the first chapter is easy to read, and it contains within it a philosophical statement that has influenced researchers in many areas of cognitive science who study things that are distantly related, at least experimentally, to vision. Among other things, the philosophical statement defines three levels of description in the study of the mind. Here are Marr's labels and tasks that Marr gives to the three levels (from Chapter 1, Figure 1-4):

• Computational theory: What is the goal of the computation, why is it appropriate, and what is the logic of the strategy by which it can be carried out?
• Representation and algorithm: How can this computational theory be implemented? In particular, what is the representation for the input and output, and what is the algorithm for the transformation?
• Hardware implementation: How can the representation and algorithm be realized physically?

At the highest level, the level of computational theory, we search for descriptions and explanations of the computational problems faced by the mind, how those problems fit into the larger cognitive picture, and what sorts of strategies the mind might use to solve those problems. This is the level at which many cognitive psychologists, linguists, and philosophers of mind operate. You look for a particular cognitive task or ability, develop hypotheses about how it works, and then look for behavioral data to support your hypotheses. At the next level, below computational theory, is the algorithmic level. Here your focus is on the specific computational properties of the cognitive system: how does it represent the information it receives and processes, and what algorithms does it use to process that information? Cognitive psychologists, linguists, computer scientists, and some neuroscientists all work at this level of description and explanation. Finally, since cognition happens in the brain, you have to understand how it happens in the brain. Thus, the hardware implementation level involves the study of the neural mechanisms underlying cognition. This is the level at which neuroscientists, along with some computer scientists, linguists, and cognitive psychologists, along with the occasional philosopher of mind work. Different individuals in different disciplines will place more emphasis on different levels of description, but Marr argues that to gain a complete picture of the mind, or vision specifically, we have to approach it from all three.

Marr does such a good job of clearly describing the importance of this approach, and the approach itself has been so influential in cognitive science, that I think anyone who is interested in the discipline should read this first chapter. Thanks to brainsci001, whose real name I do not know, for pointing me to the web copy of the chapter.