The Human Mind is a Bayes Logic Machine

lexxyz writes "Apparently the human mind can predict the distribution type for a given sample of results. A study found in The Economist has shown that a group of minds working on single pieces of data, can together generate the statistical model used to represent a given sample. Note that it takes a group of people to be able to accurately predict the behaviour of something, not a single individual"

Comments

Working "together"?

[GuyMannDude]

A study found in The Economist has shown that a group of minds working on single pieces of data, can together generate the statistical model used to represent a given sample. Note that it takes a group of people to be able to accurately predict the behaviour of something, not a single individual.

Well, that's a somehwat misleading summary. These people were not knowingly collaborating. Each person would have had to answer the questions independently (not knowing what the other respondants' answers were) in order for Bayes to be applicable. Each person's response counts as a piece of evidence or clue in inferring the underlying probability distribution. Their answers are combined using Bayes's rule by an external third party (the researchers). So, yes, this technically counts as a group of minds working together, but I think the way it this summary was worded might give people the wrong impression.

Think about it this way: if you lock a bunch of people in a room toegther and have them come up with an answer, the "strong" personalities in the room are likely to have a heavy influence on the "weaker" ones. People who aren't really firm in their opinions are going to influenced -- whether they realize it or not -- by people who sound confident. The article makes a big to-do about the fact that Bayesian techniques allow you to get good answers with a small number of people working on the problem. But the key is that those people have to be working independently because it's going to be damn difficult to identify and subtract out the cross-correlation of members influencing each other.

I'm making (what I hope to be) an important point. I think business people who read this article or even slashdotters who read the above summary may get the impression that small meetings are a great way to arrive at strikingly effective solutions. That's not what Bayes techinques are about. If you want to put a small group of people to work on a problem, you'd better separate them , otherwise Bayes's rule is not strictly applicable.

GMD

Re:Hari Seldon?

[Lon]

or "The Wisdom of Crowds" by James Surowiecki

Links

[elronxenu]

http://www.monzy.org/audience/

http://www.monzy.org/audience/ICMI-2002-finalpub.p df

http://movis.net/research/audience/techniques.pdf

http://www.cuug.ab.ca/CUUGer/9705/confernc.html

Re:Then please explain the failure of democracy

[Anonymous Coward]

For example, imagine a bayesian spam filter would come with a pre-installed distribution which says that every mail where the word Viagra appears has a 99.99999999% probability of being spam. A pharmacist would then have a very hard time to convince that filter that the Viagra orders he gets are no spam.

Not at all. Bayesian spam filters typically work by forming a weighted combination of the spam probabilities of all the words in the email, not just one word. Even if "viagra" by itself has a 99.99999999% probability of being spam, it's going to be swamped by all the (presumably) non-spammy words in the email.

Suppose the email mentions Dr. Marmaduke Smith and Marmaduke only ever occurs in non-spam emails. It won't take very long before "Marmaduke" gets assigned a 99.99999999% probability of being non-spam. An email with both "Marmaduke" and "Viagra" will then have no net probability of being spam or non-spam based on those two words alone. Now the other words in the email come into play. If there are other words like OVERNIGHT!!!!!! DELIVERY!!!! in the email, it'll look more spammy. If there are other words that normally appear in legitimate pharmacy mail, it'll look less spammy.

This is actually the fundamental power of a Bayesian spam filter over a simplistic keyword based yes/no spam filter.

Nonsense

[Anonymous Coward]

The Scientific Method requires that the hypothesis makes the predictions before they are tested.

The "Scientific Method" is a myth perpetrated by elementary school science textbooks. Actual, practicing scientists (of which I am one) do not adhere to any cookbook "method", and in particular hypotheses (let alone their predictions) do not always precede data. In fact, it is quite common for it to be the other way around, especially when you don't know much about the system being studied (exploratory data analysis) or when new statistical methodologies allow you to reanalyze data in a better way.

Now, it is important to make new predictions about data that the hypothesis wasn't fit to, but a completely different issue is being discussed here. In fact, in this case the analysis method was decided upon before the data (not that it has to be); it's just that the data collection method was screwed: it allowed respondents to give non-numerical answers ("infinity") when the analysis method required finite positive values.

For example, consider that an arbitrary amount of completely random data of just 6 variables is likely to yield at least one "statistically significant" correlation for the determined data-miner to write his paper about.

That's not because the statistical analysis method was made after the data was collected, it's because the statistical analysis method (p-values) are bogus; the inference method they're based on is not logically coherent. You can mathematically prove that the Bayesian method is coherent, and that p-values can grossly overestimate "significance".

Now, there are cases where the choice of statistical method has to be made before the data is taken (such as stopping rules in frequentist sampling theory), but those also arise because of incoherent methodology. Stopping rules, for instance, don't appear in any methodology that adheres to the likelihood principle (of which Bayesian methods are a subset).

How Does the Brain Do Plausible Reasoning?

[radtea]

One of the fundamental modern Bayesian papers is Jaynes' "How Does the Brain Do Plausible Reasoning?", which can be found on the web along with lots of other interesting things. Jaynes' conclusion is that we must be Bayesians under the skull. It's a compelling paper, even now.

These experimental results are exactly what Jaynes theory predicts, which is a very nice confirmation of his work. But they are not the "discovery" of anything--they are empirical confirmation of something we already knew. When light-bending by gravity was measured it was not a discovery, it was the confirmation of a theoretical prediction. This is the same.

Um, no. Not exactly that simple...

[Moekandu]

For a good source on the current thoughts/theories on AI try:

http://www.singinst.org/GISAI/index.html/ General Intelligence and Seed AI.
and
http://www.singinst.org/CFAI/index.html/ Creating Friendly AI.

Both really drive home the complexity of creating AI. The human brain isn't merely a "database engine that applies statistical rules to the queries it processes" . It's a carefully networked collection of highly specialized modules, of which one could be called the Bayesian Statistical Module. Bayesian statistical analysis is quite important to AI, but as Eliezer Yudkowsky (the author of the two listed papers) states, "It is necessary, but not sufficient."