Original article was published on Artificial Intelligence on Medium
Do Machines really learn?
Machine Learning and what it means
When I first heard the term ‘Machine Learning’, I wondered when did machines start learning. So did you, didn’t you? This learning machine thought takes you down the rabbit-hole of machines replacing humans, reproducing, and then conquering the world.
Good news, it isn’t that crazy … yet. Knowing what machine learning is and what it isn’t helped me discard a lot of speculations in my head. So, where do we start? Definition?
It is a field of study that gives the ability to the computer for self-learn without being explicitly programmed.
— Arthur Samuel , IBM Scientist who coined the term in 1959
This is one of the oldest definitions we know of. But, this doesn’t really help. It just takes you back to “ When did machines start learning? ”.
Let me save you from this misery. Truth be told, machines don’t learn. Let me go further, none of the machines have ever learned anything. period. Looks like we’re back to the title. But, don’t worry, let’s proceed forward.
Machine Learning is just great branding. If Machine Learning were called something like “Learning Algorithms”, it probably wouldn’t be in the hype. So, how do machines recognize images and do all the hyped learning stuff?
Here’s an example. Let’s say you wanna recognize cats. You want your machine to say whether it is a cat (1) or not a cat (0). While you can look at a koala bear and recognize them later on in pictures with 100% accuracy, machines can’t. They resort to complex mathematical computations. What a dud right?
Now, let’s see how this works. You’re wondering how can images be represented by numbers, right? After all, it’s a picture. While you look into the picture, machines look within a picture. Look at what? The RGB values.
A picture is divided into lots of pieces. These are called pixels. And the color of each pixel is determined by its RGB color values. For example, the RGB value of red is (255,0,0). This means Red is full of red (255) of course, but none of green (0) and blue (0). That way an image can be deconstructed to just numbers.
Let’s take a picture with a cat. Convert it to numbers. We then put all these numbers together to get a matrix. And by some calculations, let the machine figure out how to get 1 ( For, it’s a cat ) or close to 1 from it. This matrix is called input data. And it turns out the machine needs like 100,000 images for it to predict cats well. Even then the accuracy isn’t as good as yours. Look at you, you can identify a new animal you’ve never seen before with only one picture.
Why isn’t that learning? Because if you slightly distort the inputs, the output is very likely to become completely wrong.
*shown a cat’s picture
YOU : Cat !
Machine : Cat !
*shown same cat picture inverted
YOU : Cat !
Machine : Not really sure. Give me some time .
What I just illustrated is called Logistic Regression. One of the many algorithms for getting a machine to recognize something. If this line sounds too geeky, just let it go. You will be just fine. If not for anything else, you should be proud of yourself.