Artificial Intelligence is, at its heart, the discipline of building machines that think like humans in particular the development and display of learning and problem solving capabilities.
There are hundreds of articles out there that talk about the socio-economic impact of AI–how it will change the way in which we work and view “value” –but relatively few articles that explain some of the basic concepts in simple terms–and the crucial principles of how machines “learn”.
Whilst the big pictures and headlines are important–the ability to understand some of the science behind the data can be at once reassuring, profoundly interesting and –in some cases–terrifying.
In my case, when researching this article–I learned two things:
1. Any analogy like this can only ever skim the surface of the topic.
2. Eliezer Yudkowsky was right you conclude you understand AI at your peril!
Hot Topic: Machine Learning
Let’s get started. Machine Learning allows computers to do things without being explicitly programmed to do so. How? By learning from patterns and experience just as a child does.
Imagine a child is shown a picture of a cat and a picture of a dog (“training data sets”) by their parents (“supervised machine learning”) and every time they mistake a dog for a cat (or vice-versa), their parents correct them. Future cats and dogs that the child sees may be of different colors, shapes, breeds–however the child still knows that they are seeing either a cat or a dog because from the previous data he/she has been shown, they can determine the patterns which give the highest likelihood of that animal being a cat or a dog.
Now imagine the child’s parents simply left them with a pile of pictures of cats and dogs without explicitly telling the child which was which (“unsupervised machine learning”). As the child goes through the pictures, they could probably make a fairly good grouping between two types of animals (even though they miss the label ‘cat’ and the label ‘dog’) based upon the characteristics or patterns between the two.
Finally, imagine the child makes a series of deductive steps to determine whether the picture is a dog or a cat–and every time they reach an ultimate correct determination of whether it is a cat or a dog they receive a sweet (“reinforcement learning”). Ultimately the child would alter their way of selecting dogs from cats based upon how the outcomes of a series of decisions was rewarded or penalized.
"Machine Learning allows computers to do things without being explicitly programmed to do so. How? By learning from patterns and experience just as a child does"
Hot Topic: Deep Learning
Deep learning is a sub-set of machine learning which involves trying to get the machine to think more like the human brain in particular around the area of “feature extraction”–or in the case above, what are the features of a cat and a dog that allows an ultimate way of telling the difference between the two.
Imagine yourself now looking at pictures of cats and dogs. It’s easy to tell them apart. But have you thought about how you can do this? It’s not about number of legs, fur color, absence, or presence of fur, absence or presence of a tail. Even if you were shown part of the animal, you could probably tell the difference fairly accurately.
The reason is that there are many data points which together ultimately differentiate between whether a picture of an animal may be best classified as a dog or a cat and those data points are so wide and various that it would be impossible for a programmer to “hard-code”. Deep learning is a way of configuring a computer to identify which are the data points (or features) that need to be considered and in what way to make an accurate prediction between a dog or a cat. The word deep comes from the fact that learning happens in multiple layers, i.e. the network has depth. Each layer extracts a certain representation of the data with the ultimate goal of classifying whether an image is a cat or a dog. In our case of cats and dogs, for example, the first layer of a deep learning algorithm might extract pixel darkness, the second layer could extract the edges of the pixel darkness and the third layer would take edges and start joining them in shapes such as ears, legs and tails. Finally, the last layer would take the features from the previous layers (let’s say ears, legs and tails for simplicity) and ultimately classify the outcome as a dog or a cat.
Neural networks are often used to try to emulate the multi-layered thinking capacity of the human brain in deep learning situations– while combining it with huge computing power and speed to think “biologically” about problems.
Understanding the building blocks of AI (or at least some of them) is a first step to building a picture of how AI impacts our present and drives our future. If you made it this far, I’m grateful for your reading and any comments you can give.