Found this excellent non-technical explanation to AI and Deep Learning which I enjoyed so much I decided to paraphrase it for benefit of my own learning. The original version has more detail than mine and is presumably written by someone who unlike myself actually knows about this shit, so check it out if you want to know more.
You sit down at your new desk on your very first day, the boss comes over and puts a sheet of numbers arranged in a grid in front of you.
Boss: “Is this a cat?”
You: “Wut?”
Boss: “WRONG!” *Slaps you in the face*
Boss: *Places new sheet in front of you* “Is this a cat?”
You: “I don’t understand…?”
Boss: “WRONG!” *Slaps you in the face again and places a new sheet down*
Boss: “Is this a cat?”
You: “I don’t know?”
Boss: “WRONG!” *Another slap in the face, another sheet of numbers*
Boss: “Is this a cat?”
You: “Yes” (a complete guess because you haven’t got a clue what’s going on)
Boss: “…Yes it is a cat. Well done, here’s fifty quid”
After a full day of blindly guessing you start to notice some vague patterns; If a sheet is mostly zeros then it’s usually not a cat, and if the bottom left corner is mostly high numbers then it often is a cat. You have no idea how these patterns relate to identifying cats, all you know is they seem to work.
You hire 10 people and ask each of them if the current sheet is a cat. Your new team have a range of abilities, some are usually correct while others are always wrong, but eventually you see patterns in their answers and it helps your make your answer more accurate.
Each member of the team seems to be good at spotting certain parts of a cat: one guy might always spot eyes, then next guy can identifiy a tail, and so on. These insights aren’t much use individually but when you combine all 10 you start getting pretty good at indentifying cats.
If a team of 10 people are pretty good at identifiyng cats then what about 100 people? Or 1000 people? The more layers of people you have involved in your team the more accurate and consistent your guesses will be, but it’ll take a lot of time and energy to do all that extra work so you’ll need to find a balance between accuracy and efficiency.
Your team is working well but it takes quite long for each member to look at the entire sheet, what if you were to slice it up and give everyone a small section instead?
This idea works well, especially for those team members looking for small elements of the picture as that doesn’t require the whole sheet anyway. The technique allows the entire process to be faster and more accurate.
You and your cat identifying team are an example of a neural network. Together you make many guesses and are rewarded or punished based on if you are correct.
Nobody told you how to read the sheets or how to recognise a cat, you had to figure it our yourself based on trial and error. This is known as machine learning because the machine is learning the rules by itself.
When you hired your team you became an example of deep learning. ‘Deep’ refers to the number of levels of team members below you rather than the profoundness of the subject.
Slicing up the sheet into sections instead is an example of convolutional neural networks, they are described as a “significant step forward” in the capability of neural networks.