Computers are as Smart as a Four Year Old…What?

According to a University of Illinois, Chicago research team, a publicly available artificial intelligent computer (ConceptNet4) could produce an IQ score average equivalent to a four-year old child.  Taken out of context, this finding is astoundingly impressive, with the dawn of computer overlords just around the corner.  Applying Moore’s Law, the continually expanding processing power – and the implied commensurate advances in algorithmic control of this processing power – allows for a steady progression of higher levels of intelligence replication.  The below figure represents a typical projected computing intelligence “growth chart.”

As ever, the devil is in the details.

“’If a child had scores that varied this much, it might be a symptom that something was wrong,’ said Robert Sloan, professor and head of computer science at UIC, and one of the study’s authors.” (http://news.uic.edu/a-computer-as-smart-as-a-four-year-old).  ConceptNet4 performed relatively well on vocabulary, recall, and similarity matching.  It performed horribly on comprehension and reasoning.  A school counselor encountering a real child performing this well on recall but this poorly on comprehension may well wonder what his caregivers are doing to him at home.  Is the poor child locked up in a closet studying the dictionary all day? 

Why does ConceptNet 4, as a proxy for some of the most advanced computational approaches, produce these results?  And what are the implications for computational approaches and real world needs?  If this is the IQ of a four-year old, the question becomes “a four year old what?”   

ConceptNet, now in its version 5, is a Massachusetts Institute of Technology Media Lab sponsored approach towards making computing smart and interactive.  One can think of ConceptNet as being an academic version of IBM’s Watson.  One can also think of ConceptNet as being a massive, connected database of real-world facts and their relationships.  In fact, one of the major components of ConceptNet is a database called Open Mind Common Sense.  In the words of one of its founders,  “We have recently started a project here at MIT to try to build a computer with the basic intelligence of a person…” For more background on Open Mind Common Sense, including the motivation and history of one of the co-founders, please refer to a fascinating Wired article.   

Computers do what computers do best.  The core of any modern computer is an Arithmetic Logic Unit (ALU).  It computes.  Specifically, it performs arithmetic, logically compares, and moves data. The integrated peripheral of any modern computer is its memory – the registries, the memory chips, the hard drive/solid state drive, and any removable memory media.  These recall.  The heart of any computer program, be it a movie player, a video game, a DNA analyzer, a predictive model, or a common sense answer simulator, is to recall, add, compare, and store data according to its instructions.  A computer is a calculator. 

Ask a computer a question like, “Who was Henry the VIII?” and what it does is iteratively fetch every file in its memory database (solid state drive), compare it to the terms, “Henry the VIII,” and store matching files in another memory location for possible further processing.  Asking it to read a passage about Henry the VIII is asking it to save the file in its memory database.  These are the vocabulary, recall, and similarity matching questions.  ConceptNet is a connected database of real-world facts and their relationships.  The strength is that it makes efficient searching and connections of the open source, volunteer-added information (e.g. Wikipedia is also a major component). 

Humans do what humans do best.  If one believes a human does the same as a computer, then the child who instantly knows the vocabulary answer to any question would be the most popular child on the school playground.  The child who tells perfectly non sequitur stories would be the model child.  Then the intelligence growth chart above is relevant and will run off the scale by the end of the decade.  This would then be the day computers are our overlords. 

But if that is not what is meant by intelligence, then perhaps the child who comprehends and understands what the target classmates want and are interested in and so tailors and times the stories to amuse them wins the crowd.  If this were the case, then a memory-recall-and-comparison-centric database would only continue to produce high average IQ scores fraught with extremely high variance and imprecision.  Similarly, a computer guided by memory and pattern-centric Cerebellar Model Articulation Controllers, classic neural networks, machine learners, and especially the eidetic k-NN might be useful for high average IQ scores. They can easily show they are as smart as a four-year old.  But this would not be a four-year old human child, or chimpanzee, or cat, or mouse, or even a fish.  It would be a computer as smart as a four-year old computer.