Causality vs. correlation - statistical reasoning is not enough - NY Times Interview with Dave Ferrucci

Dave Ferrucci, who was until several months ago an IBM Fellow  and was known as the father of Watson, was interviewed by the NY Times in his new working place at Bridgewater Associates.

In the interview Ferrruci somewhat continues the line of thought of Noam Chomsky,  saying that AI has concentrated around statistical reasoning based on correlations, but the drawback is that one cannot understand why the prediction made by the statistical reasoning is correct.  While Chomsky bluntly stated that statistical reasoning does not create a solid model of the universe, Ferruci claims that a complementary approach is required -  understanding causality.    This is a rather old issue, in symbolic logic, there is a distinction between "material implication"  which states that  IF A is true then B is true, and the meaning is that always when A is true then B is also true, which makes a sentence like  "If the week has seven days than  the capital city of France is Paris" - a valid statement in logic.    Entailment, on the other hand, said that "A ENTAILS B" if it is necessary and relevant, in other word, there is a causality among them.  Thus, Ferruci concentrates now on building causality models to model the world economy.      I concur with the assertion that understanding causalities give better abilities of reasoning and prediction.   As David Luckham already noted, causality among events is one of the major abstraction of event processing models.   Here is a rather old discussion about causality of events.  

Statistical reasoning and event processing tutorial

Streambase has posted a video tutorial on combination of their EP product with statistical reasoning based on MATLAB.  The idea of combining event processing and statistical reasoning is becoming part of the big data offerings, and no wonder that statistical reasoning vendors are adding event processing to their portfolio, for example the introduction of event processing within SAS, 

Streambase takes it from the other side -event processing vendor that combines statistical reasoning.

Interesting tutorial to watch.  

More on statistical reasoning - Chomsky on "Where AI went wrong"

I have written before about the claim that statistical reasoning is over-hyped in its claim that all the problems in the universe can be solved by statistical reasoning over past data (see also my report about Sethu Raman Keynote in DEBS'12).  

A more blunt claim against statistical reasoning has been made by Noam Chomsky which claims that the fact that AI took a wrong direction by making statistical reasoning its mainstream, and claim that knowledge achieved by statistical reasoning, while having practical usages as "good enough" for various uses, creates shallow knowledge that only approximate the universe and does not create a solid model of the universe.  

An interesting interview with Noam Chomsky was published recently in "the Atlantic".    

Always thought provoking!

I'll write more about correlation vs. causality.  

How can the level of uncertainty be determined?

You may identify this formula as Heisenberg uncertainty principle,  nature is full of uncertainty and so is the world of business and any other world.  It is not difficult to convince people that uncertainty representation and handling is needed, but people wonder how in practice people will be able to consume and digest uncertainty based systems.   I'll try to refer to it in a series of posts,  the first of them deals with the following issue:  uncertainty handling models assume that the level of uncertainty can be quantified.  Quantification can be in the discrete world, a collection of values, and in the continuous world -- probability, measure of belief, and similar metrics.    The question is -- how can we determine the value that represents uncertainty.

There are three ways to determine uncertainty:   prior knowledge,  observation and  statistical reasoning.

The prior knowledge often exists due to physical properties:  the accuracy of sensor may be a property of the sensor reported by the producer, mathematical models can determine the error rate of physical measurement due to friction or other physical phenomena, this can also be rooted in statistical analysis, but for a various system it is given as prior knowledge.

Observation:  In some cases and observation includes some uncertainty, such as:  I left home somewhere between 8 - 8:30am,  there was an accident somewhere in main street that made me late for the meeting, I heard that the accident was caused due to a drunk driver (but not sure this rumor is true),  I arrived to the meeting a few minutes after the meeting start, and waited a long time to the elevator that took me into the 35th floor.     

This observation is full of inexact fact: time, space, event attributes and more.  People typically does not know how to quantify it, but can use fuzzy terms that can be translated into quantified values either in the discrete space or in the continuous space.

The statistical reasoning path is based on learning mechanism that is based on the ratio between historical input and the real value.  The assumption here is that eventually the real value is known and can be compared to the reported value.

There are some interesting questions about the representation formalism, the coverage that can be obtained by these methods, and the methodology for value assignment.    I'll write about these questions in a following post.