How would Alan Turing fix A.I.?

Computers just don't get it.

Image credit: Elliott Brown

We want speaking machines because language is the best way to rapidly communicate ideas. Users want their lives made easier and Hollywood wants their 1960s predictions proven right.

But computers don’t work for artificial intelligence (A.I.). Alan Turing, the famous British mathematician, cryptologist and a founding father of today’s computers would have pointed out that they weren't designed for it.

Computers were designed to compress and duplicate information. They don’t handle A.I. well because they leave too much work to programmers. Programmers were replaced by engineers using statistics in waves of hope from the 1970s, but the results remain inaccurate and limited. A.I. hasn't scaled and after nearly 60 years of effort, a new approach is warranted.

What?! A computer models a person?

Alan Turing’s paper from 1936 describes the computer process that is still in use today. He modeled a human who did computations with a digital computer that emulates her. Human computers at the time supported many things including ballistics, banking and science.

Both human and digital computers calculate by blindly following procedures. Interestingly, human computers were usually thought of as women. The first computers were men in the 1700s, but from the late 1800s and especially during the Second World War, computers (and the earliest programmers) were typically women.

Can you imagine it? Turing emulated the 1930s human computer that used sheets of paper to process. He did not emulate what they did to sit at the table, drink coffee, see the sheets with their eyes, think about the next step or move their hand to write numbers.

This model has held back A.I. ever since, because what a person does and how they do it are radically different things. Human speech, for example, is a totally different problem to adding 1+1 on many levels.

What else is wrong with the digital computer?

Turing defines the computer as storing a finite number of symbols -- like an alphabet and a set of punctuation and numerals. You cannot store anything else. This is the compression of information. And if you want to store the same word twice, you duplicate it by storing the same symbols twice in the same sequence. Compression and duplication are a design feature of computers and they are different to brains.

This design feature is amazingly powerful, with the Internet, iPhones, Windows and the general explosion in technology the beneficiary, but limiting for A.I.

Imagine how many times Wikipedia stores the most common English word: "the"? That is duplication on a massive scale. With A.I., the duplication creates the need for a search facility. Today, a search facility means we often guess answers because a single word can have many meanings. Worse, different words can mean the same things, such as when we store more than one language or dialect.

AI is locked out by the computer

Many in the field consider it heresy to suggest that computation is not at the core of the problems in the cognitive sciences, but that's the problem. Lack of progress means we need to consider alternatives. 

Alan Turing’s tragic death at a young age in 1954 may have cost us getting intelligent machines much earlier because we lost his visionary guidance when we needed it most. Two years later, A.I. was founded at the Dartmouth Conference. Its computer-science supporters were familiar with the details underlying computers at the time. The trouble is, those computers were based on human computers. A.I. should be modeled on human brains, machines that can handle human languages, not on what human computers did.

Where next

Science uses models to predict results. Bad models don’t accurately predict results, but good ones do. The targets that have been set in the world of A.I. aren't aligned with the expectations of customers.

Siri, Dragon, Google, Bing Translate and others illustrate the gap: users want improvements. Despite an obvious lack of progress, engineers continue to try to leverage systems based on the old models.

Imagine if normal computers were like this. If your bank account were to change from time to time due to errors, you’d be angry, while today’s free Internet translation companies hand over totally wrong translations regularly.

Turing would have told the A.I. community that computers, at their core, are not the right start for intelligent machines.

Copyright © 2015 IDG Communications, Inc.

7 inconvenient truths about the hybrid work trend
Shop Tech Products at Amazon