Saturday, August 26, 2006

Decisions, Uncertainty, and the Brain   posted by Coffee Mug @ 8/26/2006 12:19:00 AM

I just finished reading Decisions, Uncertainty, and the Brain: The Science of Neuroeconomics by Paul Glimcher though, and I would recommend it to a friend. The book is about 5 years old now, and neuroeconomics is kinda blowin' up. Glimcher believes that neuroscience has been dominated by a paradigm based on finding the minimal neural circuitry necessary to produce stereotyped reflexive behavior and that we need to shift toward understanding the goals of the nervous system to fully understand how behavior is generated. I read the book because I wanted to get a grip on the neuroeconomics and decision-making papers that are flying off the press, so the philosophical and paradigmatic issues weren't really necessary for me although they led me down some entertaining avenues of thought. If you just want an intro to neuroeconomics and the types of experiments involved, you could probably skip up to chapter 10 and read the last 100 pages or so. I think the most obvious application of this field is judicial. It seems like you could set up this problem in terms of economics and then look for neurological indicators and ?interventions?. Glimcher has addressed the law in at least one article but focused on the application of neurobiological evidence in courts. All that aside, below I just want to give you an idea of the scope of the book and the fields he draws from to make his case.

Part I: History

Chapter 1: Describes beginnings of scientific method (Francis Bacon) and application to human physiology (William Harvey). He introduces Vaucanson's duck, a mechanical duck created by a Jacque de Vaucanson which will serve as a recurring symbol for the reflex-based, deterministic view of behavior. The main purpose of the chapter is to introduce Descartes and Cartesian dualism. Descartes needed the soul to explain "unpredictable and nondeterministic behaviors...that the clockwork scientific explanations available...could not hope to mechanistically explain...".

Chapter 2: Lagrange and Laplace do so well in describing the physical universe with analytical mathematics that it seems certain that biological systems will also eventually succumb. In 1833, Marshall Hall presents a paper to the Royal Society in which he transected frog spinal cords and found reflexes in the bottom half and volition in the top (dualism creeping into physiology). There is more detail regarding the history of spinal reflex physiology the innerworkings of the Royal Society.

Chapter 3: Charles Sherrington was a smartypants, but he is this book's villain. Sherrington conceptualized behavior in terms of the reflex-arc. A receptor organ receives a sensory stimulus, this stimulation is transmitted by way of a conductor to an effector organ. This is the simplest explanation for the reflexes that we are all familiar with, but Sherrington had to bring in inhibitory components too to explain how one muscle can tighten while one loosens in a coordinated fashion. Pavlov takes reflexes up into more complex behavior, and we start to think behavior is deterministic and if we had good maths we could explain the whole damn thing. Godel looms into view to ruin everyone's party. The Godel issue seems more like zeitgeist party-pooping than a necessarily devastating issue for the reflexological paradigm.

Chapter 4: "The whole function of the nervous system can be summed up in one word, conduction." This quote will haunt Sherrington for the rest of the book. Reflex theory can't deal with spontaneous behavior (we are subjected to a thought experiment in which a cat is placed in a sensory deprivation chamber) and internally generated oscillations. We need feedback (reafference) and oscillators to explain the function of the nervous system. Also, the same goal-directed behavior can be achieved through a number of different motor programs, so the goal must be decided followed by the computation of the proper motor program, implying a hierarchy. A hierarchy is not simple conduction, so this is another challenge to the Sherrington's formulation.

Chapter 5: Computational models. Classical models have an input layer, a hidden layer, and an output layer. A speech model called NetTalk is discussed. The point is to show how much modern neurobiological models and experiments are informed by the reflex paradigm. A big chunk is devoted to explaining some interesting experiments performed by the Newsome lab demonstrating that neurons in area MT (a part of the neocortex containing neurons sensitive to moving visual stimuli) fired in direct correspondence to a monkey's reported perception of the direction of stimulus motion, and that perception could be influenced by stimulation of these neurons. Newsome employed a Michael Shadlen to model the data computationally, and they came up with a nice deterministic model to explain motion perception. Glimcher says this is all well and good, but we are limiting ourselves to unnatural experimental paradigms because of the Sherringtonian influence.

Chapter 6: David Marr is one of the heroes of the book. Marr was a theorist and computer scientist who wrote a book called Vision in the late 70s as he was dying of leukemia. Marr said you can't understand flight by studying feathers. You have to understand aerodynamics. By analogy, we can't understand the nervous system by studying neurons and mechanisms, we have to use theory and understand the goals of the brain. Marr did this for visual perception, but Glimcher notes that two immediate objections arise. One is that the definition of the scope of a nervous system goal seems arbitrary, and two is that it is not clear that evolution is concerned with computational goals or that the brain does things in an efficient goal-directed way.

Chapter 7: How to dissect up the nervous system to get some goals to understand. Fodor and colleagues argue that cognition is divided up into independent modules. Michael Gazzaniga, working with split-brain patients (much communication between left and right hemisphere is removed to isolate epileptic foci), shows that the different halves of the brain were specialized and can work independently. So cognition can be divided up into goals, and it doesn't have to be completely arbitrary. Gould and Lewontin show up to introduce phyletic and architectural constraints that could raise problems for Marr. Evolution may not be able to get you to the optimal computational solution. But Glimcher points out that there are cases where we are darn close to optimal. Rhodopsin is the molecule in your eye that converts light into a biochemical signal so it can influence neural firing. Rhodopsin can detect a single photon. That's as good as you could hope for. There is also a section on convergent evolution in African cichlids suggesting that an evolutionary goal may be important than how you get there.

Part II: Theory and experiments

Chapter 8: What's the overall goal? Maximizing inclusive fitness. "I want to suggest that we can characterize the function of the nervous system as decision making." Decision making should produce behavior that increases inclusive fitness. It is easy to make optimal decisions if you are omniscient, but organisms living in the world have to make decisions in the face of uncertainty. We need probability theory to deal with uncertainty, and we ought to define the goals of the nervous system in terms of probability theory. A history of probability theory featuring Blaise Pascal follows. Probability of an event X value of that event = Expected Value. Value doesn't predict how people will behave though. One man's trash is another's treasure and all that. Bernoulli came up with Expected Utility to describe the case in which, for instance, the same absolute amount of food (Value) is more motivating when you're hungry than when you're satiated. We get a fairly detailed description of Bayesian likelihood estimation. Bayesian theory is the optimal tool for decision-making in uncertain conditions. Economics uses these tools to predict behavior. We should use economics to define the goals of the nervous system and call it Neuroeconomics.

Chapter 9: John Krebs' behavioral ecology defines optimal goals using economics and explains natural animal behavior in terms of decision-making. The "prey model" is developed in this context and uses probability theory and economic models to define the most efficient predator behavior. When should a predator bother to collect food and when should it pass? The prey model is empirically tested, and a bird called the titmouse seems shows behavior very close to that predicted by a probabilistic economic model. This shows the utility of defining behavioral goals using economic models.

Chapter 10: This is the longest chapter and the meat of the book. We get an introduction to the neuroanatomy of both visual processing and motor control of the eye. Primates can be trained to move their eyes based on visual stimuli. We know how the info gets in and where it goes out, but where is the sensory input connected to motor output? Research has focused on an area that seems anatomically well-positioned for the job called the lateral intraparietal area (area LIP). A major leap occurs when Herbert Jasper and Edward Evarts develop a system for recording from single neurons in awake behaving primates. A controversy ensues in which one camp insists that parietal neurons encode high-level motor commands and another camp says they are rather involved in high-level visual processing (i.e. attention). There are a lot of experiments described with regard to this controversy. I think Glimcher is trying to illustrate how bogged down we can get when trying to interpret neuro-data using the receptor-conductor-effector model.

Glimcher and Michael Platt stepped into the fray in 1997, but their major move came in 1999 when they tested the idea that the LIP was encoding economic parameters necessary for making decisions about which way to move the eyes. Neurons in the LIP have a preferred position in the visual field that they would like to fire for. Glimcher and Platt varied the probability that this location would be the target location (the place to which the monkey must move its eyes to receive a reward) across trial blocks. The firing rate of neurons in the LIP corresponded directly with the prior probability of reward at their preferred location. When, conversely, the probability of reward was held constant and the amount of reward was varied, neurons in the LIP also showed a direct relationship between activity and reward value. Probability of reward X value of a reward = Expected Value (and perhaps Expected Utility, see here for more).

In the prior experiments, the monkey was eventually told by a visual cue which way to look for reward, but Glimcher wanted to ask whether monkeys make probabilistic decisions in the face of a probabilistic problem. He drew on an observation by Richard Herrnstein that pigeons could match their behavior to an experimentally controlled probability landscape, so-called Matching Behavior, to produce a paradigm in which monkeys could demonstrate their matching proficiency (even though this wasn't the optimal strategy). Monkeys did model the probability and value setup with their actions and neurons in the LIP tracked the estimated value of a given eye movement. Glimcher isn't satisfied though because the monkeys didn't perform the economically-defined optimal deterministic strategy but instead behaved probabilistically. The next chapter attempts to explain why sometimes probabilistic behavior actually is the optimal course predicted by game theory.

Chapter 11: Intro to Game Theory. This is also my first real encounter with game theory, so expect me to be getting it wrong for a while. Starts with Von Neumann and Morgenstern. The optimal solution when faced with an intelligent adversary is sometimes to become as unpredictable as possible. When your best move is to vary which type of move you make you have arrived at a mixed strategy solution. You act unpredictably. VN and M only did the math to predict mixed strategy solutions in competitions in which one entity's gains are mirrored by the adversary's losses (zero-sum). John Nash was able to expand the math to include non-zero-sum games. The treatment of Nash's contribution is fairly detailed. Glimcher says that Nash allowed us to consider that nondeterministic (uncertain, probabilistic) behavior is often optimal. I'm still not clear on why this is Nash's insight instead of VN and M's, but I will take Glimcher's word for now. John Maynard Smith brings game theory into biology and evolutionary theory, and finds probabilistic behavior is evolutionarily stable. Finishes with a description of experiments by D.G.C. Harper showing that ducks behave according to the laws of economic theory. Glimcher makes kind of a big deal out of the idea that animal behavior can be 'fundamentally unpredictable' and 'irreducibly probabilistic'. I'm thinking that his usage of the term irreducible is a little uncommon. In the next chapter he attempts to discover neural correlates of this unpredictable property. This seems like a step toward reduction of the problem to me, such that if we really knew the brain, the behavior would no longer be uncertain.

Chapter 12: The work or shirk game is introduced pitting a worker against a boss checking up on him. People act according to the principles discussed above and act really randomly. Glimcher and Mike Dorris took the game to a computer versus a monkey. Monkeys act right too. Neurons in the LIP fire at a consistent rate when probability and value of rewards are varied but relative expected utility is maintained. If you look at a finer grain, you can see that neurons are adjusting their firing rates with updates after each trial and that these updates work in the manner predicted by a Nash equilibrium solution.

Chapter 13: This is a wrap-up chapter. Glimcher proposes a bunch of other experiments he's done or would like to do. He has a couple concerning visual attention and motion. Some on the neurobiology of encoding prior probabilities. He mentions the basal ganglia as a potential spot for this type of learning, but doesn't really review what the basal ganglia is/are all about. There is also some on the encoding of value and how dopamine might be involved.

Chapter 14: This is the second wrap-up chapter focused on philosophy. Descartes was wrong. We don't need dualism. We don't need reflexes at all. We need a continuum of behavior from probabilistic to deterministic. He seems to have a pretty nice take on free will. "Free will may simply be the name we give to the probabilistic behaviors that are mixed strategy solutions. Our subjective experience when a mixed strategy solution requires the activation of a lawful neuronal randomizer." Return to Vaucanson's duck. Animals act indeterminately aaaand they don't have to have souls to do it.

Hope this gets some people interested in reading the book and maybe getting further into the area. I am going to be trying to discuss some of this stuff as it comes out. In general, I get more excited about brain mechanisms than ecology or theory, but these perspectives seem to have been really important in developing Glimcher's views. I could've stood for a little more neuroscience background. We got a fairly good treatment of the LIP, but I'm really not familiar with single-neuron recording, and I'm sure there are caveats and pitfalls to this research technique (one being that it is purely correlational). One obvious next move is to either attempt to manipulate expected utility behavior by stimulating this area or to affect the same by ablating the area.