By Pat Dixon, PE, PMP
President of DPAS, (DPAS-INC.com)
I recently read "The Sentient Machine" by Amir Husain. It was written in 2017, so it may be considered outdated since artificial intelligence (AI) is a very rapidly progressing technology. However, the book is still a worthy read.
A considerable portion of the book deals with the potential dangers of AI that have been expressed by people like Geoffrey Hinton, Bill Gates, Stephen Hawking, and Elon Musk. In the middle position was Marvin Minskey, one of the pioneers in AI, who acknowledged some risks but was less alarmist. On the other side of the spectrum are people like Ray Kurzweil who see the tremendous benefit and brave new world that AI can build.
Like Kurzweil, Husain sees a relatively rosy future for AI. In his book he explains the many potential applications of AI that can provide benefits unachievable otherwise. Husain begins with a chapter entitled "What is AI?", which is necessary to explain terms before talking about them.
Husain states "Artificial Intelligence is a broad field of study". Broad means that it is more important to define what is NOT AI than what is. So much of what we use computers for can be considered AI. Husain says that machines that can "reason, learn, and act intelligently" are application of AI. Consider a software application for the game of chess. The application has the rules of the game embedded as well as the optimization algorithm for determining the best move. Such an application is "hard coded" with these rules and algorithms, but under Husain's definition is considered AI. Chess games have been in computers since the first consumer market for them appeared. You may not have thought the Commodore 64, Apple Lisa, IBM PC Jr, or Tandy 1000 had AI, but Husain says it did under a broad definition.
Much of what we consider AI today is a subset called Machine Learning (ML). ML uses data to train algorithms. In contrast to hard coded logic, algorithms are built to learn logic from data. ML is where most of the recent advances in AI have come from.
What is NOT AI? Humans obviously. Artificial means non-human intelligence, so we are not AI. That means anything else that is not intelligent is not AI, and intelligence was described by Husain as "reason, learn, and act intelligently". This is obviously circular logic. We haven't really defined intelligence.
Obviously, a hammer is not reasoning, learning, or acting. Neither was a Ford Model T. However, today's cars can sense the road and detect when you are out of your lane to automatically steer you back between the lines. Is a dog AI? The bigger challenge is to explain what AI is not and why.
A common method for discriminating AI is the Turing test. While Alan Turing, the great mathematician who invented this test, apparently meant it as a joke, it might be the best approach for defining AI. Put a human in a room where they don't know what is outside the room. They must interact with something outside the room. If the human can't tell whether they are interacting with a human or an intelligent machine, the intelligent machine is AI.
Consider a thermostat. Your interaction is to give it a temperature setpoint. A human outside the room could look at that setpoint, compare it to a thermometer, and manually decide to turn on or off a heating or cooling unit. This would be tedious, unrewarding work and very inefficient use of human labor. The thermostat does the same work more effectively and efficiently, but the human inside the room wouldn't know the difference. Is a thermostat AI?
In manufacturing, we have the equivalent of thousands of thermostats which are implemented as PID loops, which are the foundation of supervisory and advanced automation strategies built on top. We have had this in industry since the 3rd industrial era arrived about 50 years ago, and under Husain's definition would be considered AI. ML is in its infancy in manufacturing but is maturing. Usually when we talk about AI in manufacturing, we really mean ML.
Considering the dangers of AI applied to manufacturing, we are very far away from the nightmares that some fear. I don't know of any AI application in industry without humans in the loop. Our data has noise and outliers. Training ML models without humans to pre-process data with filters and outlier detection is very dangerous. Any closed loop application in industry required human supervision.
I recommend "The Sentient Machine". It is relatively concise, well written, and pertinent work for any of us that interact with AI.
