Slashdot Mirror
Physicist Proposes New Way To Think About Intelligence
An anonymous reader writes "A single equation grounded in basic physics principles could describe intelligence and stimulate new insights in fields as diverse as finance and robotics, according to new research, reports Inside Science. Recent work in cosmology has suggested that universes that produce more entropy (or disorder) over their lifetimes tend to have more favorable properties for the existence of intelligent beings such as ourselves. A new study (pdf) in the journal Physical Review Letters led by Harvard and MIT physicist Alex Wissner-Gross suggests that this tentative connection between entropy production and intelligence may in fact go far deeper. In the new study, Dr. Wissner-Gross shows that remarkably sophisticated human-like "cognitive" behaviors such as upright walking, tool use, and even social cooperation (video) spontaneously result from a newly identified thermodynamic process that maximizes entropy production over periods of time much shorter than universe lifetimes, suggesting a potential cosmology-inspired path towards general artificial intelligence."
http://www.alexwg.org/publications/PhysRevLett_110-168702.pdf
Yes, what you're missing is the entire point of the paper. Here's my attempt at a quick summary:
Suppose you are a hungry crow. You see a tasty morsel of food in a hollow log (that you can't directly reach), and a long stick on the ground. The paper poses an answer to the question: what general mechanism would let you "figure out" how to get the food?
Many cognitive models might approach this by assuming the crow has a big table of "knowledge" that it can logically manipulate to deduce an answer: "stick can reach food from entrance to log," "I can get stick if I go over there," "I can move stick to entrance of log," => "I can reach food." This paper, however, proposes a much more general and simple model: the crow lives by the rule "I'll do whatever will maximize the number of different world states my world can be in 5 seconds from now." By this principle, the crow can reach a lot more states if it can move the stick (instead of the fewer states where the stick just sits in the same place on the ground), so it heads over towards the stick. Now it can reach a lot more states if it pokes the food out of the hole with the stick, so it does. And now, it can eat the tasty food.
The paper shows a few different examples where the single "maximize available future states" principle allows toy models to "solve" various problems and exhibit behavior associated with "cognition." This provides a very general mechanism for cognition driving a wide variety of behaviors, that doesn't require the thinking critter to have a giant "knowledge bank" from which to calculate complicated chains of logic before acting.
Oh you're a fan of webcomics? Here, have a Penny Arcade.
Here's a review of this paper by a researcher who actually works in the field of AI and cognitive psychology:
Interdisciplinitis: Do entropic forces cause adaptive behavior?
Few choice quotes:
and after he explains what the paper's about and how utterly empty it is, he offers some advice to authors:
Suggesting that the purpose of intelligence in this man's random musings might be to increase the background levels of entropy for your own benefit.
That's close, I think. I am not a physicist and I skimmed the equations, but here's my take on what they're proposing. Physical systems have states, which can be described by a state vector. The state of these systems evolves according to some set of rules that describes how the state vector changes over time. They've built a simulator in which the probability of a certain state transition is computed by looking at how many different paths (in state space, i.e. future histories of the system) are possible from the new state, in such a way that the system tries to maximise the number of possibilities for the future. In one example, they have a particle that moves towards the centre of a box, because from there it can move in more directions than when it's close to a wall.
They then set up two simple models mimicking two basic intelligence tests, and find that their simulator solves them correctly. One is a cart with a pendulum suspended from it, which the system moves into an upright position because from there it's easiest (cheapest energetically, I gather) to reach any other given state. The other is an animal intelligence test, in which an animal is given some food in a space too small for it to reach, and a tool with which the food can be extracted. In their simulation, the "food" is indeed successfully moved out of the enclosed space, because it's easier to do various things with an object when it's close compared to when it's in a box. However, in neither case does the algorithm "know" the goal of the exercise. So they've shown that they've invented a search algorithm that can solve two particular problems, problems which are often considered tests of intelligence, without knowing the goal.
Then, they use this to support the hypothesis that intelligence essentially means maximising future possibilities. Another way of saying this, I think, is that an intelligent creature will seek to maximise the amount of power it has over its environment, and they've translated that concept into the language of physics. That's an intriguing concept, relating to the concept of liberty, power struggles between people at all scale levels, scientific and technological progress, and so on. I can't imagine this idea being new though. So it all hinges on to what extent this simulation adds anything new to that discussion.
On the face of it, not much. You might as well say that they've found two tests for which the solution happens to coincide with the state that maximises the number of possible future histories. The only surprising thing then is that their stochastically-greedy search algorithm (actually, without having looked at the details, I wouldn't be surprised if it turned out to be yet another variation of Metropolis-Hastings with a particular objective function) finds the global solution without getting stuck in a local minimum, which could be entirely down to coincidence. It's easy to think of another problem that their algorithm won't solve, for example if the goal would be to put the "food" into the box, rather than taking it out. Their algorithm will never do that, because that would increase the future effort necessary to do something with it. Of course, you might consider that pretty intelligent, and many young humans would certainly agree, although their parents might not. It would be interesting to see how many boxed objects you need before the algorithm considers it more efficient to leave them neatly packaged rather than randomly strewn about the floor, if that happens at all.
There's another issue in that the examples are laughably simple. While standing upright allows you to do more different things, no one spends their lives standing up, because it costs more energy to do that as a consequence of all sorts of random disturbances in the environment. The model ignores this completely. Similarly, you could
How is this paper not a scientific approach to empirical data? We have empirical observations of a wide range of animal/human behaviors. The authors propose a toy mathematical model that reproduces key features of several interesting observed behaviors. This is perfectly good science, just like saying "hey, an F=G*m_1*m_2/r^2 force between massive objects recreates the observed motions of the heavenly bodies" --- a predictive, testable mathematical model that can be compared with measurements of the motions and behaviors of actual critters to see how well it works.