Saturday, April 27, 2019

Ockham's Guillotine: Minimizing the Argument Surface of the Social Sciences

Established interests view "Trumpism" as foreboding a modern "storming of the Bastille".  They wonder, "What form of guillotine will populists roll out this time?  Will my head lop in a bucket?"

This article suggests rolling out an inherently judicious "guillotine":

Ockham's Guillotine

Ockham's Guillotine lops only the metaphorical heads of the social pseudoscience Bastille.  It precisely guides Ockham's Razor down the grooves spanning only those necks.

This it does by selecting the best unified model of society, based on a single number:

The model's size, measured in bits of information.

Deprived of "wiggle room" for their "lies, damn lies and statistics", social pseudoscientists will be helpless as the Razor slips, the buckets receive and the crowds roar.

The science for Ockham's Guillotine is here; its mechanisms driven by one of the most powerful forces in history:

The explosion of computation and social data detonated by Moore's Law.

Harnessing this raw power in the design of Ockham's Guillotine requires a theory equal to the task:

Algorithmic Information Theory

Algorithmic Information Theory (AIT) is the computational form of Ockham's Razor:

The Algorithmic Information content of any data, including social data, is the size of the smallest program (algorithm) that outputs that data.  That program necessarily embodies the smallest model of the data.

AIT founds the exploding field of Artificial General Intelligence (AGI).  AIT is sometimes called "Algorithmic Probability Theory" or "Solomonoff Induction".

The final advice given by seminal artificial intelligence figure, the late Marvin Minsky:
"The most important discovery since Godel is Algorithmic Probability which is a fundamental new theory of how to make predictions given a collection of experiences... This is a beautiful theory... that will make better predictions than anything we have today and everybody should learn all about that and spend the rest of their lives working on it."
By "experiences" Minsky meant observations/measurements in the form of data, such as social measurements.

Minimizing the Argument Surface

In cybersecurity, the concept of "attack surface" is the number of ways that an external actor can interact with the system.  Each way in which there is an interaction presents a potential vulnerability to attack.  These interactions can be viewed as "dialogues" based on the communications protocols so, in that sense, they can also be viewed as "issues" over which "arguments" can obtain.

Human conversations between potential adversaries are similar in that the more ways an issue presents itself, the more ways in which sophistic arguments can exploit the social contract upon which civil discourse obtains.  The social sciences are particularly problematic in this respect, involving a myriad of "issues" over which arguments may turn into sophistic exploits.

Ockham's Guillotine reduces the argument surface of the social sciences to just two issues:

1) Is the basis of the artificial intelligence industry valid?
2) What data is relevant to social policy decisions?

#1 is supported by the most powerful modern force as previously described:  The explosion of data, computation and economic incentives behind AGI.

#2 can be dealt with by the simple expedient of not arguing about it:  If the sophists demand that their data be included in the corpus, then let it be included.  This is enabled by the explosion of computation capacity on the one hand, and the ruthless nature of AIT on the other.  AIT is ruthless in the sense that any biased data will best be modeled by algorithmic explication of said bias that corrects the data accordingly.  Then the corrected data will be brought into consilience with the larger body of knowledge/data rather than standing alone, which requires more bits of information.

Perhaps the most ruthless approach to "nuking the social pseudosciences" would be a monetary prize for improvements in the unified model of society.  An exemplar of this kind of prize is The Hutter Prize for Lossless Compression of Human Knowledge, which targets natural language modeling based on Wikipedia's corpus.  Prize awards are paid out for each incremental improvement in the compression of Wikipedia.

A prize of this sort, targeting a unified model of society could trigger an avalanche of activity resulting from the social pseudoscientists attempting to take on the juggernaut of industrial artificial intelligence at the same time that hundreds or thousands of young people, eager to prove their chops (and earn money) increasingly embarrass the corrupt authorities of academia.

A preliminary data set of a wide range of longitudinal social measures for Ockham's Guillotine is available as an example at github.

2 comments:

scottynx said...

In the non-pc sphere, lots of us thought that Artificial Intelligence would eventually spew out non-pc results, embarrassing the pc-elites and corroborating our view of reality. But then in the 2010s we started seeing the tech industry, press, politicians and acadamia devote considerable effort towards promoting the enshrinement PC sensibilities in AI in response to early non-pc results from algorithms that were written to dispassionately tease out causal factors. As Eric Weinstein puts it, google biases search results in a pc way when it allegedly "unbiases" them. The plan is to do this to all AI.

Reading your blog entry provides a little hope, despite it being a bit outside my comprehension. Basically, to simplify for myself, I take it you are proposing that AGI will naturally have an iterative process that corrects the orwellian "unbiasing" to model reality in a way that gives predictions comporting with reality.

Jim Bowery said...

scottynx said: "AGI will naturally have an iterative process that corrects the orwellian "unbiasing" to model reality in a way that gives predictions comporting with reality."

More precisely, "Iterative approximation of algorithmic information content of a sufficiently rich corpus naturally explains (models) bias in the corpus as meta information."

This iterative approximation will most likely be driven by human intelligence, not AGI, for the same reason as are optimal data compression algorithms. The relationship to AGI is primarily in that AGI, the exemplar for which is Hutter's AIXI, is a unification of Algorithmic Information Theory with Sequential Decision Theory: Algorithmic Information embodies the best model of what "is" (pure science) and Sequential Decision Theory embodies the best model of what "ought" to be (pure engineering specification).