[Music] [Applause] Thank you, thank you, thank you. So, during World War II, the first computer was invented; it cracked a German communication code and ensured a successful Normandy landing. The father behind this unprecedented machine, Alan Turing, wrote the paper "Computing Machinery and Intelligence" in 1950.
The paper opens with the words, "I propose to consider the question: Can machines think? " Well, today, inspired by his thoughtful question, we'll try to answer the following: How can we create an intelligent computer, and what will the future look like with intelligent machines? Well, in fact, AI has been growing exponentially in the past decade.
It has already been touching our lives in ways that you might not notice. For example, every time you go on Google search, some kind of AI is being used to show you the best results. Every time you ask Siri your question, natural language processing and speech recognition are being used.
So, artificial intelligence will probably be one of the biggest scientific breakthroughs in the 21st century. It will give us the power to probe the universe and our humanity with a different approach. AI has the potential to forever change our humanity.
The backbone of artificial intelligence is machine learning, and I think the term is pretty self-explanatory. We want to make machines learn based on their knowledge and make decisions. Machine learning can be understood in two major components.
One is to use algorithms to find meaning in random and unstructured data, and the second part is to use learning algorithms to find relationships between that knowledge and improve that learning process. So, the overall goal for machine learning is actually quite simple: it is to improve the machine's performance on certain tasks. That task can be predicting the stock market or complicated ones such as translating articles between languages.
The screenshot that you see right now is actually a depiction of Google Translate's neural network. So, speaking of translation, anyone here speak a second or third language? Great, that's awesome!
Well, I was born in China, and I speak Chinese and also speak English, plus a couple of programming languages if you want to count that. So, when my family and I travel around the world, we often need something called Google Translate. By examining Google Translate's artificial intelligence, we can actually gain a great understanding of how most AI works.
Well, first of all, have you ever wondered how much data Google has? Well, it turns out Google holds right around 10 to 15 exabytes of data. Well, what does that even mean?
Let me put that into perspective for you: If one personal computer has 500 gigabytes, then Google's 15 exabytes would be equivalent to 30 million personal computers. Data turns out to be one of the fuels that powers Google Translate, this magical technology. So, on the surface, Google Translate hasn't changed since 2007, when it first launched, but what you might notice is that the translations are getting faster and more accurate.
So, it turns out the learning process for Google Translate is inspired by our own. We, as humans, get better at doing things by practicing, just like what our math teachers and our music teachers always tell us. It turns out Google Translate can get better at translating by reading more articles.
So, how do computers learn? We can actually come up with this flowchart that will summarize and give us a good picture of how artificial intelligence actually works. It turns out we have to use some training input and put that into a learning algorithm, which will give us some knowledge.
That knowledge will be what a computer knows about that specific subject. And you and I, we—the user—right? The user will give the computer some input, and hopefully, some output will come out.
So, in our case, Google's 15 exabytes of data will be the training input, and something you want to translate is going to be the user input, with the output being something in a different language. So, the most important part of this whole entire process is actually the learning algorithm. This is what powers computers to learn and be intelligent.
So, today, we're going to focus on two parts: one is image processing, and the second part is neural networks. So, let's begin by talking about image processing. We cannot talk about computer vision without talking about human vision, right?
The visual signal from our retina is relayed through our brain to our primary visual cortex, located at the back of our brain, which is right here. Visual information is separated and processed in three different systems: one system mainly processes information about color, the second one about shape, and the third one about movement, location, and organization. With all of that in mind, today we'll try to create an application that will be able to identify a Coca-Cola logo.
First of all, we have to understand that most pictures we see on a computer screen are made of pixels—tiny things that represent color. This is also why Steve Jobs named his company Pixar, since every person in that world is made of pixels, which is great. So, when the computer is trying to understand this image, we will first separate it into different features—objects that we can easily see in the still image.
Each of these features will provide the computer with some information about that image. Today, we'll mainly focus on area, parameters, and skeletons, along with some details about these features. Now that the computer has those things in memory, when the user gives the computer some input, it will be able to process that input, compare it with what is in memory, and then give you some output, whether the image matches the template or not.
So here's that technology in action. I've created an application on this iPad that will be able to identify the Coca-Cola logo, and this application is actually powered by Open Computer Vision, thanks to a great framework. So today, we'll learn about the Coca-Cola logo.
Let's click on that. Great! We just learned this image—wonderful!
As you can see, the image on top has little green rectangles and squares around it, and those are regions the computer is processing. In the image below is one of the biggest features in that image, and in the table, as you can see, there are details the computer just remembered. So let's dismiss that and click "Start Tracking.
" Oh, look at that! It was pretty sensitive; it successfully told me that the paper right in front of me has a Coca-Cola logo on it—great! And also, this is live, so you know I'm not faking anything, by the way.
So wonderful! Thank you. Now, let's summarize everything we did with this simple flowchart.
We had some input data, and we used some algorithms to find meaning in that data. In the future, we'll use neural networks to improve this entire process and hopefully learn more and more about images. The pixels, in our case, or the input data and the meaning were things like area, parameters, skeletons—those, you know, details the computer focused on.
Hopefully, in the future, we'll be able to classify any image we want. Remember, in the very beginning, we talked about how there are two parts to learning algorithms, right? The second part is neural networks, so let's talk about that a little bit.
Our brain is made of gazillions and gazillions of neurons, and those tiny things communicate with each other to process information, and that's how we become intelligent. It took thousands and thousands of years of evolution, and it's such an amazing process! So, T has thought, what would happen if we actually turned that and put it into a computer?
First of all, we have to understand the differences and similarities between an artificial neuron and a biological one. On your left, this is a biological neuron, and it has cell bodies, axons, terminal axons, dendrites, and stuff like that. Those parts will take in information, process it, and give you some output.
Similarly, on our right, as you can see, we have a bunch of x's. From our algebra class, you might know that x represents inputs in our case, and f(x) is a mathematical calculation, while y is an output. So this picture represents the relationship between neurons since we have so many of them, right?
By altering the relationships between our neurons—which are called synapses—we will be able to learn and gain a better understanding of things. Synapses are represented as lines on our right. This is an animated version of what scientists believe our brain would look like.
Back in the old days, you know, in the 1970s and before most of us were born, when scientists wanted to do something like image recognition or speech recognition, they had to sit around a table and put papers and pens down and start doing math. They had to create lookup tables, and this was a pain because it took so much manpower and a long time. Scientists wondered, what would happen if we gave the computers the power to learn?
That would be magical! Because lookup tables would never exist if we could just make computers learn on their own. Instead, we would have computers with their own knowledge about a specific subject.
This is what this diagram represents: the computer's own knowledge about something. This is really empowering because scientists no longer have to create lookup tables for days and years. What they have to do is just write a simple program, train the computer, and then it can do things like image recognition and speech recognition in a matter of seconds.
With help from Google Cloud Platform, we're going to do another demonstration showing the power of combining image processing with neural networks. Once again, this is all live! We have a great audience here tonight, and we're going to take a picture—let's say a picture of my phone—to see what the computer thinks.
Oh, it's a mobile phone! It's a product; it's a gadget! That's wonderful!
Now, what if we take a picture of the audience? It's a performance audience! Say hi to the camera—great!
Thank you! All of the things we just talked about are intangible, just like art, music, and language. Yet technology like this plays such an important role in our daily lives.
For example, in Google’s self-driving car project, they use image processing to identify the difference between a police vehicle and a normal passenger car. This is another picture from Google’s self-driving car project. They combine image processing with laser and ultrasonic sensors to form three-dimensional models of the car's surroundings, allowing the car to navigate safely without lag.
This might surprise you: back in the '90s, scientists actually implemented this technology on fishermen's boats. A well-trained computer can identify the difference between a tuna and a cod! So next time when the dining hall serves you fish, you might appreciate the technical journey that the little fish took to be on your plate.
So what's next? Let's try to answer this question: what will the future look like with AI? Well, actually, let's jump back in history and talk about one of the biggest breakthroughs that we had with AI.
Of you might recall this historical event between Garrick Alboro and the IBM computer Deep Blue. The IBM computer became the first-ever program to defeat a World Chess Champion under tournament rules in a classic game. It was a very significant victory; it was a milestone.
However, later analysis actually played down the intellectual value of chess as a game that can be simply defeated by brute force, which means that if you had enough calculation and enough computing power, chess can be defeated. This means that calculation does not equal intelligence, and this is a very important understanding. However, Google took a different approach.
They created AlphaGo, a program that can learn the game of Go. I mean, no pun intended there, but Go is a game with far fewer rules but requires far more intuition. You cannot just calculate the possibilities of Go.
So, Google's AlphaGo was able to defeat the South Korean Go champion Lee Sedol in a 2016 game, and this was another breakthrough because the program used reinforcement learning as well as neural networks, which resemble our own decision-making process. So, AI will not only change our lives in small ways, like we talked about above; it will likely bring us tremendous change, like we saw 200 years ago with the Industrial Revolution when humans first harnessed the power of coal and steam engines. Change like we saw in the 1990s when millions and millions of computers reached homes across the globe.
AI will give us an unprecedented amount of power as well as the opportunity to change. Imagine—imagine 10 years from now when we're autonomously constructing a space station on Mars. Your car is driving you to work while you are talking to a friend on the phone who works on Wall Street, and he doesn't have to worry about stock cheaters anymore because AI will ensure a fair and safe trading environment.
Also, in hospitals across the globe, scientists are using AI to find mutations in human DNA databases and also cures for diseases. These are just some of the possibilities, and the sky is no longer the limit. The power and the freedom that we have with artificial intelligence are empowering but also humbling.
We, as humans, are capable of creating machines that can learn and think just like us. In the long run, AI will not replace biological intelligence; yet, it will enhance our lives. It will enhance our future, and I believe that most AI researchers out there will agree with me on that.
So, after all, you and I and all of us are on this journey together. All of us have the chance to witness and also decide how artificial intelligence will shape our future. Thank you.