intelligence is a large umbrella. Under it, you'll find visual recognition (“That's a cat!”), voice recognition (you can say things like, “It won't turn on” or “It won't connect to the Internet” or “It never arrived”), natural language processing (“I think you said you wanted me to open the garage door and warm up your car. Is that right?”), expert systems (“Based on its behavior, I am 98.3 % confident that is a cat”), affective computing (“I see cats make you happy”), and robotics (I'm acting like a cat).
THE MACHINE THAT LEARNS
The magic of machine learning is that it was designed to learn, not to follow strict rules. This is the most fundamental aspect to understand and the most important to remember when you hit that inevitable frustration when things start going slightly off‐track. A rules‐based system does exactly what it's told and nothing more. We are comforted by that. A command to send out a gazillion e‐mails with the “<first_name>” after the salutation does precisely that. That's good.
Of course, when the database has something fishy in the first_name field, then somebody gets an e‐mail that begins, “Hello, Null, how are you?” Once humans know to look for those sorts of mistakes, we create processes to check and correct the data before hitting Send the next time.
When a batch of e‐mails goes out that all say, “Hello, <first_name>, how are you?” and the e‐mails all include those brackets and that underline, we know to flail the programmers until they find the errant semicolon that caused the problem. In both cases, we can backtrack, find the problem, and fix it.
Machine learning has data and is given a task. It performs the task and waits for feedback. If the feedback is positive, it tries to duplicate the same actions the next time for a similar task. If the feedback is negative, it requires information on where it went wrong.
Guess the Animal
In 1980, at the Personal Electronics store in Santa Barbara, California, I was shown a program that could guess the animal you had in mind. I was flabbergasted. It was a simple case of Asimov's “Any sufficiently advanced technology is indistinguishable from magic.”
It: Think of an animal…
Is it an elephant? (y or n)
Me: n
It: You win. Help me learn from my mistake before you go … What animal were you thinking of?
Me: a rabbit
It: Give me a question to distinguish a rabbit from an elephant.
Me: Is it a small animal?
It: For a rabbit, what is the answer to your question? (y or n)
Me: y
It: Thanks.
Play again? (y or n)
Me: y
It: Think of an animal …
Is it a small animal? (y or n)
Me: y
It: Is it a rabbit? (y or n)
Me: n
It: You win. Help me learn from my mistake before you go … What animal were you thinking of?
Me: a frog
It: Give me a question to distinguish a frog from rabbit and an elephant.
Me: Is it green?
It: Thanks.
Play again? (y or n)
Me: No, I see how this works now.
It: I don't understand.
Play again? (y or n)
Me: n
It: Goodbye
Is this program rudimentary? You bet. Was this machine learning? Almost.
After running again and again, the game could guess exactly what animal you had in mind after only a few questions. It was impressive, but it was just following programmed logic. It was not learning. Guess the Animal could update its rules‐based database and appear to be getting smarter as it went along, but it did not change how it made decisions.
The Machine that Programs Itself
Machine learning systems look for patterns and try to make sense of them. It all starts with the question: What problem are you trying to solve?
Let's say you want the machine to recognize a picture of a cat. Feed it all the pictures of cats you can get your hands on and tell it, “These are cats.” The machine looks through all of them, looking for patterns. It sees that cats have fur, pointy ears, tails, and so on, and waits for you to ask a question.
“How many paws does a cat have?”
“On average, 3.24.”
That's a good, solid answer from a regular database. It looks at all the photos, adds up the paws, and divides by the number of pictures.
But a machine learning system is designed to learn. When you tell the machine that most cats have four paws, it can “realize” that it cannot see all of the paws. So when you ask,
“How many ears does a cat have?”
“No more than two.”
the machine has learned something from its experience with paws and can apply that learning to counting ears.
The magic of machine learning is building systems that build themselves. We teach the machine to learn how to learn. We build systems that can write their own algorithms, their own architecture. Rather than learn more information, they are able to change their minds about the data they acquire. They alter the way they perceive. They learn.
The code is unreadable to humans. The machine writes its own code. You can't fix it; you can only try to correct its behavior.
It's troublesome that we cannot backtrack and find out where a machine learning system went off the rails if things come out wrong. That makes us decidedly uncomfortable. It is also likely to be illegal, especially in Europe.
“The EU General Data Protection Regulation (GDPR) is the most important change in data privacy regulation in 20 years” says the homepage of the EU GDPR Portal.11 Article 5, Principles Relating to Personal Data Processing, starts right out with:
Personal Data must be:
* processed lawfully, fairly, and in a manner transparent to the data subject
* collected for specified, explicit purposes and only those purposes
* limited to the minimum amount of personal data necessary for a given situation
* accurate and where necessary, up to date
* kept in a form that permits identification of the data subject for only as long as is necessary, with the only exceptions being statistical or scientific
EU GDPR Portal, http://www.eugdpr.org.
