Gary Marcus has a marvelous article in this week’s New Yorker that points out the over-reach the media has imposed on Neuroscience. Neuroscience has progressed, but is still in its infancy. I have no disagreement with Marcus. This post is part of a conversation.
A major culprit has been functional imaging (fMRI) as presented in the media. Functional imaging has been a wonderful advance that permits a glimpse of the activity in regions of the normal human brain. A wonderful tool, but it is a tool with serious limitations. It doesn’t lead to direct understanding how the brain works. Let me explain.
Lets imagine you are interested in cognitive function X, perhaps reading. Before doing any research, its not stretch to say that the information processing that is involved in reading goes on in the brain. Somehow, the visual input from the eyes is processed in brain and reaches a stage of “comprehension”. That there is understanding is determined by the ability of the subject to answer questions on the reading material. Now you put this individual in an fMRI machine and do the same thing. The individual reads and understands. During reading the regions A, B and C are selectively activated. What have you learned? At best, you have transferred the function of reading from a big box (brain) to a smaller box (regions A, B and C). But you have learned nothing about how visual stimuli were converted into meaning. You have not learned a thing about how brain, or regions A, B and C, processed information. Moving a question from a big box to a small box is not an explanation.
To learn processing you must understand the nature of the algorithm processed in a particular region. This means understanding the neural circuitry that produces an input-output function. There is no region of cortex where this is well understood.
A second argument is that A, B and C are not ‘reading centers’, they are algorithm processing units. They know not what they do. I think of a brain region like a pocket calculator. Put in a number, press the square root key and get a result. This procedure will perform the same function whether the process is used to find the area of a square or a standard deviation. Regions A, B and C are blind processors. Were we to do fMRI on this calculator we might say that the square root key was a area processor or a standard deviation processor, but this is misleading. Neurons only know their inputs and outputs (if that). The square root key has no concept of area. Its only when we view the brain as a whole we can say the a function is performed.
Let me give an example from the region I study: the hippocampus. Many neurons in the hippocampus behave as ‘place cells’, with a single place cell firing when the animal (rat) crosses a small region of space. It has been proposed that, collectively, the hippocampus is a cognitive mapping system essential for efficient navigation. I believe all of this. But the single place cell knows nothing of place. At best, it knows when its neuronal inputs fire. We call them place cells because we see the regularity of cell firing with observations of the rat’s location in space. We (I) guess that somehow the hippocampus, or brain, as a whole can integrate this information. But single place cells cannot. Furthermore, although lots is known about the circuitry of the hippocampus, we don’t understand the algorithms that create place cells. We are making progress, but we’re a long way off.
A third point is that it is likely these regional algorithm processors are likely similar to each other. Region A may be slightly faster than region B, and region B may have a larger input set than region A. Or region A may have more direct input from vision than region B, etc. As a general rule, when a a small region is inactivated or damaged, especially in a child, the function is not totally, permanently lost. This suggests that secondary regions can perform a slightly inferior version of the algorithm and subsume the ‘lost’ function. How functional allocation is assigned remains a mystery.
In brief, functional imaging is a wonderful tool. It tells us where to look to investigate the computational algorithms. But it tells us nothing about the deep problems, what the algorithms are and how they operate, in series and in parallel, to perform complex cognitive tasks.