Inductive Learning

For a very long time, Artificial Intelligence (AI) seemed no match for the supposedly “easy” tasks. “Easy” as in the ones that even small children can do, like lifting blocks and stacking them to form a tower. Or identifying objects in a pic. The problem wasn’t just about processing power or good enough sensors, because those were getting better and cheaper all the time. So what then was the problem?

But before I get to that question, remember “machine learning”? Over-simplified, it is the most popular way machines learn:

1)      Load just a very few, very high-level rules into the computer;

2)     Throw a whole lot of data at the computer to interpret on its own;

3)     Let the computer create its own rules, include changing existing ones;

4)     Repeat steps 2 and 3.

Step #3 is why it is called “machine learning”. It is also the attempt at AI that has yielded the most results from voice assistants on smartphones to tagging photos by their content.

Was the problem that we don’t (consciously) know how we do a lot of mundane things, so we couldn’t program the steps into a machine? A la “The Puzzled Centipede”:

A centipede was happy quite,

Until a frog in fun,

Said, “Pray tell which leg comes after which?”

This raised her mind to such a pitch,

She lay distracted in a ditch,

Not knowing how to run.

Is that why machine learning worked? Because we didn’t try to teach the computer how to walk and just let it figure it out on its own by throwing data at it? Isn’t that the same way a child learns the grammar of a language, not by giving it specific rules (which we don’t know anyway), but by trial and error and observations?

Kathryn Schulz wrote in her book, Being Wrong:

“This strategy of guessing based on past experience is known as inductive reasoning.”

But it comes with its risk:

“It means that our beliefs are not necessarily true. Instead, they are probabilistically true… You make best guesses based on your cumulative exposure to the evidence every day.”

I guess this explains why Google’s photo recognition software got so good (it followed a very human-like inductive reasoning). Unfortunately, it also led to the misidentification of pics of black people as those of gorillas. Does that mean the biases of humans are based on a flaw in the way we learn?