Glimcher: Foundations of Neuroeconomics Analysis

Over the last several weeks, I have been slowly assimilating Paul Glimcher’s Foundations of Neuroeconomics Analysis. Most of the neuroeconomics that I had read previously was written by economists (particularly behavioural economomists) who have ventured into neuroscience. Glimcher is a neuroscientist who has ventured into psychology and economics. It appears to me that this make a very profound difference.

First of all, neuroscientists (and biologists in general) treat the human brain (and more generally the animal brain) with an enormous amount of respect. The biologists’ view is that an organ that has evolved over hundreds of millions of years must be pretty close to perfection. For example, Glimcher points out that the ability of a rod cell in the human eye to detect a single photon of light “places human vision at the physical limits of light sensitivity imposed by quantum physics” (page 145). Similarly, the detection of image features in the visual cortex uses Gabor functions which also have well known optimality properties (page 237).

This view needs to be reconciled with the findings of psychologists and behavioural economists that the human brain makes the most egregious mistakes on very simple verbal problems. Glimcher provides one answer – evolution performs a constrained optimization in which greater accuracy has to be constantly balanced against greater computational costs (the brain consumes a disproportionate amount of energy despite its small size). Once again, this trade-off is carried out in a near perfect manner (pages 276-278). I would think that Gigerenzer’s Rationality for Mortals is another way of looking at this puzzle – many of these verbal problems are totally different from the problems that the brain has encountered during millions of years of evolution.

The second profound difference is that biologists do not put human behaviour on a totally different pedestal from animal behaviour. They tend to believe that the neural processes of a rhesus monkey are very similar to that of human beings. After all, they are separated by a mere 25 million years of evolution (page 169). Economists and psychologists probably have a much more anthropocentric view of the world. On this, I am with the biologists; in the whole of human history, anthropocentrism has at almost all times and in almost all contexts been a delusion.

This leads to a third big difference in neuroeconomics itself. Much of Glimcher’s book is based on studies of single neurons or multiple neurons and is therefore extremely precise and detailed. Highly intrusive single neuron studies are obviously much easier to do on animals than on human beings. Much of the neuroeconomics written by economists is therefore based on functional magnetic resonance imaging (functional MRI or fMRI) which provides only a very coarse grained picture of what is going on inside the brain but is easy to do on human beings. The problem is that if one reads only the fMRI based neuroeconomics, one gets the feeling that neuroscience is highly speculative and imprecise.

Glimcher’s book also leads to a view of economics in which economic constructs like utility and maximization are reified in the form of physical representations inside the brain. I am tempted to call this Platonic economics (drawing an analogy with Platonic realism in philosophy), but Glimcher refers to this as “because models” instead of “as if models” – individuals do not act as if they maximize expected utility; they actually compute expected utility and maximize it. There are neural processes that actually encode expected utility and there are neural processes that actually compute the argmax of a function.

One of the interesting aspects of this process of reification is the detailed discussion of the neural mechanisms behind the “reference point” of prospect theory. Glimcher argues that “all sensory encoding is reference dependent: nowhere in the nervous system are the objective values of consumable rewards encoded. ” Glimsher raises the tantalising possibility that temporal difference learning models could allow the reference point to be unambiguously identified (page 321 et seq).

Another important observation is that directly experienced probability and verbally communicated probabilities are totally different things. When random events are directly experienced, there are neural mechanisms that compute the expected utility directly without probabilities and utilities being separately available for subsequent processing. As predicted by learning theory, these probabilities reflect an underweighting of low-probability events (because of a high learning rate). Symbolically communicated probabilities are a different thing altogether where we find the standard Kahnemann-Tversky phenomenon of overweighting of low-probability events.

Expected subjective values constructed from highly symbolic information are an evolutionarily new event, although they are also hugely important features of our human economies, and it may be the novelty of this kind of expectation that is problematic. ... If [symbolically communicated probabilities] is a phenomenon that lies outside the range of human maximization behavior, then we may need to rethink key elements of the neoclassical program. (page 373)

This too is probably related to Gigerenzer’s finding that frequencies work much better than probabilities in symbolically communicated problems and that single event probabilities are handled very badly.

Posted at 8:09 pm IST on Wed, 14 Mar 2012         permanent link

Comments