As of late computer scientists have gotten a new shiny tool. This new tool can take any form and can solve almost any problem without having to actually fully understand it. This awesome new tool is called machine learning. Although many people in the media and elsewhere have branded it as AI, I plan to show you that it is only a mathematical trick and nothing else.
To show you what I mean, I’ll have you do machine learning in less than 1 minute. Yes, you read that right. If you follow the next paragraph you will be able to do machine learning in less than 1 minute.
So first open up a spreadsheet. Now go to this other spreadsheet and copy all the data over. As you can see, the columns contain data from MLS on housing prizes. What we are going to do is predict the price based on the house size. So go ahead and select the size and price and price columns. Now go to Insert at the top then click on Chart. A new window pops up. Go to customization and scroll all the way to the bottom. Where it says trendline, click and select linear. Below it select equation instead of custom. If you do this right you should see a graph of the prices with a straight diagonal going through it.
And voilà, pat yourself on the shoulder since you just did machine learning. You can now predict the price of any house based on its size. Using the equation on the trendline you can predict the price of any house. For example if you have a new house which has size of 5000 squared feet, you just have to do some basic maths:
97.12 * 5000 + 82124.09 = $567724.09
Using your machine learning system you would predict that the house should be valued at $567724.09.
And that is machine learning at its core. It’s all about grabbing a function and juggling its parameters until it represents best the data. Of course the simple linear function has a lot of error. A two-parameter function would never be able to represent the complexities behind house pricing.
However what if we use more variables? If you go again to the second spreadsheet to the Solved tab you’ll see a second graph with a more complicated function. It now also uses the number of bathrooms and bedrooms to estimate the value with a total of 4 parameters. You can see how the forecast is slightly closer to the actual data (See here if you are interested in how I got this function).
What computer scientist do is create arbitrarily complex functions which they call neural networks, and use fancy algorithms to modify the parameters so that the function progressively gets better at forecasting the output data. Just to give you an idea, while our linear function here had two parameters to wiggle around, Google uses around 5,000,000 parameters for image recognition. Yes that’s 5 million numbers that, if set right, can predict whether the image contains a bird among other things. And of course Google uses billions of sample data instead of just hundreds as we did.
And now coming back to the main topic of discussion, hopefully you can now see why these types of algorithms can never be conscious in the way portrayed in science fiction. Machine learning is only about learning a function that matches the data. This is why I agree with Andrew Ng when he says Fearing a rise of killer robots is like worrying about overpopulation on Mars. Thinking that machine learning will bring about strong AI is similar to people in the 1950’s worrying that the then new computers would become conscious.
Author: Andres Paez
Tech’er, writer, programmer, life enthusiast.