Understand these 5 basic concepts to sound like a machine learning expert
Aaron Edell is co-founder of Machine Box — a machine learning startup designed to make it easy to start building things with machine learning.
Most people seem a bit intimidated or confused by machine learning. What is it? Where is it going? Can I have some money now please?
All valid questions. The truth is, you’ve been training machine learning models for years now, probably without realizing it. Do you use an iPhone or Apple photos? Or how about Facebook? You know how it shows you a group of faces and asks you to identify them? Well, by tagging those photos, you are training a facial recognition model to identify new faces. Congratulations, you can now say you have experience training machine learning models! But before you do, read these machine learning basics so you can accurately answer any follow up question.
1) The benefit of machine learning is that it can predict
If you’re just tagging your friend’s faces in pictures, you’re not using a machine learning model. If you upload a new photo and suddenly it tells you who each person is, then you’re cooking with machine learning gas. The whole point of machine learning is to predict things based on patterns and other factors it has been trained with. It can be anything; housing prices based on zip code and number of bedrooms, likelihood of a flight delay based on time of year and weather, tagging of objects or people in pictures etc.
2) Machine learning requires training
You have to tell a machine learning model what its trying to predict. Think about how a human child learns. The first time they see a banana, they have no idea what it is. You then tell them it is a banana. The next time they see one (not the one you trained them on because you already ate it) they’ll identify it as a banana. Machine learning works in a similar way. You show it as many pictures of a banana as you possibly can, tell it its a banana, and then test it with a picture of a banana it wasn’t trained on. This is an over simplification a bit because I’m leaving out the part where you also have to tell it what isn’t a banana, and show it different kinds of bananas, different colors, pictures from different perspectives and angles etc.
3) 80% accuracy is considered a success
We are not at the point in technology where a machine learning platform will achieve 100% accuracy with identifying bananas in pictures. But that is ok. It turns out that humans aren’t 100% accurate either. The unspoken rule in the industry is that a model with 80% accuracy is a success. If you think about how useful it is to identify 800,000 images correctly in your collection, whilst MAYBE not getting 200,000 correct, you’re still saving yourself 80% of your time. That is huge from a value perspective. If I could wave a magic wand and increase your productivity that much, you’d give me lots of money. Well, it turns out I can, using machine learning, so please send check or cash.
UPDATE for 2018: The 80% rule is now more like the 90% rule.
4) Machine learning is different from AI, deep learning, or neural networks
People tend to throw all of these terms around casually. To sound like an expert, learn the difference.
AI — Artificial Intelligence just means a computer that is as good as (or better than) humans at doing specific tasks. It can also mean a robot that can make decisions based on lots of input, not unlike the Terminator or C3PO. Its a very broad term that isn’t very useful.
ML — Machine learning is a method for achieving AI. It means making a prediction about something based on training from sets of parsed data. There are lots of different ways a ML platform can implement training sets to predict things.
NL — Neural networks is one of these ways a machine learning model can predict things. Neural networks work a bit like your brain, by tuning itself through lots and lots of training to understand what a banana is supposed to look like. You create layers of nodes that get very deep.
5) We have a ways to go before AI becomes self aware
I’m not worried about robots taking over there Earth just yet. Mostly because if you’ve ever built a machine learning model, you know how much it relies upon you as a human to tell it exactly what to do. And even when you give it clear instructions, it usually gets it wrong. You have to be so explicit with these systems that the chance of it suddenly becoming sentient is remote. Even a simple web page that shows a box with a word in it requires you to tell it exactly where that box appears, what shape it is, what color it is, how to work on different browsers, how to be displayed correctly on different devices etc. etc. etc.
There is much standing in the way of even a very deep neural network taking over the world and turning us into batteries, mostly because no one told it to do that (hopefully).
