By Cupid Chan, CTO, Index Analytics
I recently took my kids to Hersey’s Park in Pennsylvania. In case you haven’t heard about it, it’s just a normal attraction park with rides, and long lines. As we were waiting in line, my son asked, “Dad, what are you doing at work?”
I said, “I help my clients to define KPIs, and then try to apply Naive Bayes to predict the outcome. If the result is not good, we may need to build a neural network, and test it again.”
Do you really think that’s the answer I gave my son?
OF COURSE NOT!
Not because what I said is wrong, but he is simply not the right audience for that type of response. More importantly, I don’t want him to think “My dad is crazy and I’d better not ask him anything again.” So, I need to come up with an answer in a language that he can understand.
“If a computer can do work but no one knows whether it’s you doing the work or the computer, that’s AI.” – a basic principle of AI proposed by Alan Turing.
“Great! I can then use AI to do my homework and my teacher would not know that it’s not me doing that!”
“Hmm… Do you remember how you taught your younger sister the difference between a pen and an apple? You hold up a pen in front of her so she can see it and say, ‘pen.’ And you hold up an apple so she can see it and say, ‘apple.’ And you repeat this. Sooner or later, you expect her to understand the long pointy thing is a pen. And the red, round thing is an apple.”
Long, pointed, round, red. These are Features in Machine Learning. And “Pen” or “Apple” are Labels. Combined, this is Supervised Learning. This is one way how a computer can understand that different Features are associated with different Labels in Supervised Learning.
“Dad, I remember I saw a guy teaching people this on YouTube, too!”
Well, the song is funny but it is not related to Supervised Learning. But if it inputs the concept of Supervised Learning for a child, why not let it be?
In the real world, Supervised Learning can help in many different ways. One of them is distinguishing between a cancer cell from a normal cell. In this case, the computer is the “child” and the doctor is the “parent.” By showing examples repeatedly, the doctor trains the computer to distinguish the patterns between a normal cell and a cancer cell.
You may have heard about the Law of Entropy, or the Second Law of Thermodynamics. In general, unless you put in energy to keep the situation in that current state, the whole condition will just become messier over time.
You can apply the very same law to a kid’s playground. Unless you really put in effort to keep toys tidy, the toys will not automatically go back to their original positions. At my home, my mother-in-law helps out the kids to keep the play areas organized. Once, when she went to Hong Kong for a vacation, the play areas became more disorganized day after day. Finally, my wife had to step in and demand that the kids clean up before grandmother returned. She did not give exact instructions. She just demanded they clean up!
Guess what happened in the next few hours? The kids put all the four-wheels-boxy-shaped things in one area, and we called it “Cars.” And all the fluffy stuff was put together in another area, and we called it “Stuffed Animals.” And then they put all the blocks that can be stacked up together in some boxes and named “Legos.”
They did not get any specific instructions or rules to decide what should go where. But somehow they figured out the similarities and differences. In Machine Learning, this is called Unsupervised Learning.
This is when the computer is given a lot of data points and the computer figures out the pattern by itself. In the real world, Unsupervised Learning can be used in customer segmentation. There is a lot of information and data about a lot of customers. You don’t tell the computer who should be grouped with whom, but this is figured out by Unsupervised Learning. Traditionally, this is done by the expert who observes different patterns, like age, spending pattern, where you live, salary… and then tries to group the types of customers together. And now, we have the machine to play the role of expert, which is able to scan through millions of records in a few seconds but is impossible for any human being
When dealing with kids, it’s not always the best way to just keep telling them and keep showing them the proper examples. At the same time, it’s not very effective to give no instructions and let them figure out everything by themselves.
It’s a common practice in teaching kids to reward them when they do something good. And when they do something bad, you punish them. This is intended to reinforce certain behaviors. In Machine Learning, this is known as Reinforcement Learning.
When a computer performs the way that you want, you add a point. When it fails to do what you want, you reduce a point. The computer therefore knows what to do to gain points.
In the real world, Reinforcement Learning is applied heavily in Robotics. For example, a robot is trying to walk a straight line. It may make it or it may fall down. Whenever the robot falls down, you reduce a point. And whenever the robot successfully makes one step, you add one point. There are many motors and sensors on a robot, and all of them are collecting data for the system. The robot learns what kind of motor speed, what kind of angle is needed in order to keep walking in a straight line and avoid falling.
2 Types of Measurement
2 Popular Questions by Kids – Key Approaches in Machine Learning
Kids like to ask a strangers, “How old are you?” and “Are you a boy or a girl?”
“How old are you?” is asking for a number. It’s Regression.
“Are you a boy or a girl?” is Classification. Looking for an outcome for a pre-defined category. Both are 2 important concepts in Machine Learning.
3 Ways to Learn
Kids observe the world around them. They come up with certain rules. They will propose the result, and they will be corrected by adults. Which makes the rule to get better and better.
Compared to the old way of programming: Developer observes the world. They code rules using rule-based algorithms. And they will come up with some results. Based on this, they will change or modify the rules.
In AI, it’s a little bit different. Developer creates the AI algorithm and have it create the rule. The algorithm comes up with a model and continue to train it. The model then tries to predict the result and see if it is accurate or not. The key here is that the algorithm keeps modifying the model using more data without the developer being involved.
That’s the beauty of AI!
No Right or Wrong. Just Right or Left!
Final question: What are the similarities and differences between Tesla and Uber? They both are both in the automobile industry. But one company, Tesla, creates new technology to help revolutionize the whole car industry. While Uber uses existing technology (like mapping, mobile app..etc) to create a new business model.
So the power of AI is not just in making algorithms. It can be using existing algorithms to build new ways of doing business. One builds the technology, one utilizes it.
Remember my son who was thinking about ways to get his homework done? Ultimately, I would be equally proud if he came up with an algorithm that could do his homework and successfully fool his teacher or if he utilized existing algorithms to do the same thing. Both are important new ways of adopting AI to solve problems.
There is no Right or Wrong, only Right or Left. But no matter which direction you pick, be persistent and you will cross the finish line of success via either route – Cupid Chan tweet on Nov 28, 2018
The content of this blog has been presented in a few national and international conferences such as Open Source Summit in Shanghai China and MicroStrategy Federal Summit in Washington DC. I also captured this in my very first YouTube channel video which you can find here: https://www.youtube.com/watch?v=dh9xz4SBukE&t=13s