These posts are largely unedited “brain dumps” which I dicated into my phone. I am publishing them for those who might find them useful.
There was this fight in the 1970s between Noam Chomsky and B. F. Skinner. Or, rather, Chomsky had been attacking Skinner and behaviorism. It went so far that Skinner, in his last public speech before his death, warned of the dangers of “structuralism”.
Basically, the different positions were these:
Skinner said that psychology is effective when we drop the notion of a “mind”, and focus instead on that which we can see. That is: What an organism does, and what the circumstances are under which it does it. This was Skinner’s antidote to the endless probing into the “soul” of people in the Freudian style, which never seemed to accomplish anything.
Skinner’s method, on the other hand, succeeded not only in teaching pigeons to play ping-pong or steering rockets. It also succeeded in building teaching machines for kids, and was the foundation for clicker training, Direct Instruction and the Michel Thomas Method of ultra-fast language learning.
Skinner went so far as to say that even human speech was based on conditioned behavior.
This is where Noam Chomsky came in. He attacked Skinner’s book “Verbal Behavior”. He argued that humans started out with an innate inner structure, a “language instinct”, if you will. Only because of these innate structures, or templates, Chomsky says, was a child able to learn language at all.
Chomsky’s influence grew. Skinner died in 1990. And today, the history books basically tell us that Skinner was a freak who raised his own daughter in a “Skinner box”, and that all the good in psychology and in human-computer-interactions were the fruits of the “new” dicipline of “cognitive psychology”. (Even M. Mitchell Waldrop’s excellent book “The Dream Machine” makes that mistake.)
I think, if you look closely, you see that the real winner… is Skinner.
Skinner maintained that there was no such thing as and inner template. He conceded that you do start with an organism which has its own specific possibilities and limitations.
A chimpanzee, for example will never be able to learn to speak because of, among other things, the way its vocal cords are arranged, as famously discovered by Michael Jackson in the 1980s.
But that, otherwise, the mind functions purely through reinforcement learning.
And I believe, history has proven Skinner right.
First of all, pretty much all the things we are seeing on the Internet, pay per click advertising, optimization algorithms, game design, Facebook, have their foundation in behaviorism.
The ideas of Chomsky, on the other hand, have remained pretty much theoretical.
And when we look at learning, specifically, we can now see much better how it works:
We have neural networks that are learning language, that are learning to drive on roads designed for human drivers. And what we are seeing it that this is all due to neural networks learning, without any prior innate structure.
There is no structure of the language inside the neural network before it starts to learn. Instead, that structure is in the language itself. It is outside of the learning organism. Namely in the form of reinforcement or the lack there of. It is the job of the neural network to, through trial and error, map out this landscape.
Skinner was completely right, when he said, that people usually misinterpret classical conditioning. People like to say that Pavlov’s Dogs learned to “associate” a ringing bell with food. But Skinner maintained:
The dog didn’t associate these two things with each other. Pavlov did.
It is the same with machine learning today:
It is not the computer that makes associations. Instead, it merely recognizes and maps associations which exist in the outside world.
When Tesla’s Super-computer learns how to drive a car, it receives two inputs. One input contains the images from the car’s cameras. And the other input are the labels. There are the labels provided by Tesla’s data scientists, which say “this picture contains a bicycle”, “this picture contains a traffic light”, and so on. And labels generated by the fleet, in the sense that a driver might turn the steering wheel to the left in a certain environment, or override Autopilot when it makes a mistake.
It is the neural network that creates, out of this mess of right and wrong, an “inner structure”.
And the interesting thing is, that we do not care how that inner structure looks like. We cannot look inside the neural network. And we do not need to! We don’t need to go all Freudian on our neural networks.
All that we care about is the behavior of the network. All that we care about is that the network performs its task with an acceptably low rate of error.
That is exactly what Skinner was talking about all along.