What is artificial intelligence, deep learning, machine learning?

Artificial intelligence (artificial intelligence - AI) now appears everywhere. It's something used to reply automatically to Gmail on email, learn how to drive us to play, rearrange photos of outings into individual albums, and even help manage the house well. shopping anymore. But do you know that artificial intelligence is not simply an entity, but is it divided into smaller categories? What are the current limitations of artificial intelligence products? And why don't we need (or don't need) to worry about artificial intelligence flaring over the world?

What is neural network, machine learning, deep learning?

Neural networks are not necessarily hardware based, they can still be software and algorithms.

These are phrases that you often see in information related to artificial intelligence. Basically, you can think of these things as layers of artificial intelligence.

Neural network, temporarily called artificial neural network , will lie at the bottom. This is a system of computers and computer system devices that are generally connected in some way to simulate a part of how neurons work in the human brain. The computers in the neural network can be located close to each other in the same room or thousands of miles away from each other, each of which can be viewed as a nerve unit, called a node .

Neural networks are not necessarily hardware based, they can still be software and algorithms.

The concept of neural network dates back to the 1950s with the advent of artificial intelligence research. It is said that when located separately, these computer nodes only run what is pre-programmed and can only answer simple questions, or in other words, it is "not smart". Just like in the human body, a neuron has not made a difference, but when connecting them into a thick network, things will be much different. When computer systems are connected, they can solve more difficult problems. And most importantly, when applying the right algorithms, people can "teach" computers.

The next class is machine learning. This is a program that runs on the neural network, it will train the computer to "learn" something, such as learning from the line of users' handwriting to guess if it is What characters, or learn from thousands of beach photos to find common ground and then look at the flyer to know this plate right away, not the mountain.

Deep learning is located at the top. This is a special branch of machine learning science. Deep learning has become popular in the last decade thanks to the rapid increase in the amount of digital data that mankind has created, in addition to the increased processing power of computers while costs have fallen. Will talk more about deep learning below.

How does artificial intelligence work?

Suppose you want a computer that knows how to cross the street. In the traditional way, we will program it to look left, right, how to wait for the car to run through, how to follow the right track according to law and many other things, then let the machine go by itself.

As for artificial intelligence, specifically a machine learning program, you will let your computer see 10,000 videos of how to safely tape. Next you show him 10,000 more videos, but this time it shows people getting hit by a car when they hit the road. Now you let it freeze itself.

Deep learning uses multiple layers in a neural network to analyze data in many different aspects.

The hardest part is that you have to make the computer understand and acquire information from these videos, just as the most difficult part of teaching is to make students understand what you say and remember it. . For decades, many different methods have been used to teach computer learning. One of the ways that is used is "reinforcement learning" , which means you will "reward" the computer when it does exactly what you want and then slowly optimizes for the best results. People often train animals in this way. Another way is "natural selection" , which means that many ways to solve the same problem will be applied in parallel, this is the fastest and most accurate solution to win the remaining ones.

In this day and age, people use a method called deep learning. Deep learning uses multiple layers in a neural network to analyze data in many different aspects. For example, if you give a computer a picture of a deep learning technique, each layer in this artificial neural network will see the problem in a unique way. The bottom layer will simply draw a 5x5 grid into eight images and mark "yes" or "no" when an object appears in the cell. If "yes" , the upper layer will begin to look at each of these cells more carefully, it analyzes whether this is the beginning of a straight line, or is this an angle? So many layers will help the software understand complex issues, all based on breaking it down and "investigating" slowly. It is also for this reason that people call this "deep" , ie deep and multi-layered.

The technique is applied by Facebook to identify faces, which divide the picture into different layers to learn.

Yann LeCun, head of Facebook's artificial intelligence division, said: "When you go to higher classes, the things that are discovered will get bigger and bigger. More and more aspects are being analyzed. And when you climb to the highest grade, you will have tools to tell you that picture is taking people or a little dog or an airplane ".

We have only recently talked about getting to know, now it's time to teach the computer what it has just realized. A neural network is used again, but this time it will consider many characteristics of a cat. Many pictures of cats will also be shown to the system along with the following statement: this is a picture of that cat. Then they showed another series of pictures of dogs, pigs, bears and ducks and told them: this is not a cat. Through such a series of data, the software will know what cats usually have in common, how their nails, feathers, limbs, heads and tails are called cats .

Over time, the machine will remember these data and arrange them in order of importance. For example, claws are not only available for cats, but if the nails come with big feet and mustache, this is exactly the cat. Relationships like these will also be provided from time to time during the process of learning machine software learning images. This process takes a long time and is repeated many times. Every next time it will be better than last time because of human suggestions or even other artificial intelligence systems.

You may find that just to be aware of a cat is too complicated, while Facebook, Google or Microsoft machine learning systems must recognize many other things in life. So, Microsoft's pride in releasing an application that can identify difficult fast breeds sounds simple but behind it is a complex artificial neural network and has begun to learn. in a very long time.

Images created from Google Drea artificial intelligence project.

Is this something that Google, Facebook and many other companies are using?

The general answer would be "That's right".

Deep learning is currently being used for many different tasks. Large technology companies often set up a division of artificial intelligence. Google, Facebook also open their software learning systems for all users. Google last month also opened a 3-month course on machine learning and deep learning.

Some examples of machine learning are the Google Photos tool. It has the ability to identify and categorize images that you take on different topics, even by different faces automatically. Or Facebook M, a personal virtual assistant, half-person and half-machine can help you put some items you want. Microsoft with Cortana, Google with Google Now and Apple with Siri are all very real examples of artificial intelligence.

This is also the reason why in recent times we have begun to hear more about machine learning, deep learning and artificial intelligence. That's because big consumer companies have started to jump into the game and offer real products that people can grasp and experience. Earlier, deep learning and machine learning were only in laboratories at research institutes and universities.

Some other applications of deep learning

Suggestion system on platforms

Suggestion system helps increase user interaction.

Large platforms such as Facebook, Amazon, Netflix, . all have a very strong (recommend) suggestion system that significantly increases user interaction. Specifically, they are based on user data generated when suggesting additional products they will like (on shopping platforms), the movies they will want to watch (eg, on Netflix), suggest advertising / sponsored articles (on Facebook) or interested learner courses (on online learning platforms).

Image recognition

The goal of image recognition technology is to identify and identify objects in the image as well as understand the content and context in it. The example above shows AlchemyVision's face recognition and identification service that distinguishes two similar faces between actor Will Ferrell and drummer of the Red Hot Chili Peppers, Chad Smith. Image recognition technology is also included on Facebook to suggest users tag their friends' faces or apply to crime science and investigate.

Detecting rare diseases

IBM Watson's artificial intelligence has discovered a kind of disease that doctors have given up.

Recently, IBM Watson's artificial intelligence discovered a disease that bundled doctors could not find in a female patient. By comparing this woman's genome to more than 20 million other research results, Watson has shown an extremely rare leukemia in just 10 minutes.

Limitations of current artificial intelligence

Deep learning is being used for things like speech recognition and image recognition, things that are potentially commercially viable. But in parallel, it also has many limitations.

First , deep learning needs a huge amount of input data so that computers can learn. This process takes a lot of time, a lot of processing power that only large servers can do. If there is not enough input data, or there is enough data but not enough power to handle, everything cannot happen as intended, the result of the machine learning will be inaccurate.

Second , deep learning is still not able to recognize complex things, such as common relationships. They will also have trouble recognizing similar things. The reason is because there is currently no technique good enough for artificial intelligence to draw those conclusions logically. Besides, there are still many challenges in integrating abstract knowledge into machine learning systems, such as information about what that object is, what it is used for, how people use it . In other words, machine learning does not have the same knowledge as humans.

A very specific example for you to understand: in a Google project, a neural network is used to create a picture of a dumbbell that people often hold in the gym. The results are quite impressive: two gray circles are connected by a horizontal tube. But there is a human arm in the middle of this tube, and this is not something in the "problem". The reason is quite easy to guess: the system is taught about dumbbells with photos of people being weightlifting, so it is obvious that sticking one's hands. The system can know how dumb, but it does not know that the fruit will never have an arm.

Besides bringing many benefits, deep learning is also limited.

With some simpler pictures, mahcine is still confused. Experiment by a group of researchers showed that when they showed the computer a series of images that had only random pixels, they were certainly 95% a . truck, or a starfish. .

That's not all. According to computer scientist Hector Levesque, the current artificial intelligence tools use many "tricks" to erase the true gaps in their knowledge. Virtual assistants like Siri or Cortana often make you feel like you are talking to real people because they use jokes, quotes, emotional expressions and many other things, just for you to divide center.

Try asking for things that need to be thoughtful, such as "can a bad fish drive?" Or "a soccer player is allowed to mount wings to fly or not". Questions of this type are too complicated for today's artificial intelligence systems, so often no results will be returned to you, if any are also irrelevant or simply search the sentence of you on the internet.

Winter AI

The artificial intelligence industry is an easy industry but easy to get down. In 1958, the New York Times talked about a machine that distinguishes left and right as a kind of intelligent robot. But so far, we have not been able to create a robot with such intelligence. And when those promises are not made, people use the word "winter AI". That is the period when the amount of money invested in AI has plummeted, few have mentioned it, and people are also skeptical about the possible results. So far there have been about 6 small "AI winters" and 2 big seasons that appeared, in the late 70s and early 90s.

Perceptive artificial intelligence?

Many people working in artificial intelligence think that it will be very difficult for us to create a sentient artificial intelligence."There is very little evidence at the present time that shows the hope of creating a highly flexible artificial intelligence and doing things that they were not created to do," said Professor Andrei. Barbu from MIT. He emphasized that the study of artificial intelligence today only creates optimized systems to solve a specific problem.

There have also been some research work on computer learning but without supervision, ie just give data to machine learning without labeling right or wrong or explaining anything. However, Andrei Barbu commented that projects like this have not progressed and are still far away to reach the results date. An example that once appeared was a Google neural network system that randomly took a thumbnail of 10 million videos on YouTube to teach itself what the cat looked like. However, Google said this was just an experiment and said nothing about its accuracy.

In other words, we still do not know how to make computers self-learn without supervision. That is the biggest barrier. That is, it is still far away from the day when robots can sense and fight people.

As Elon Musk said, his company created an artificial intelligence for self-driving on Tesla cars. But he never said that he would know everything. This is simply a network that helps cars learn from each other. When one learns something, others know the same thing. The end result is not cars that can do everything in the world, it's just to solve a very specific problem.