Artificial intelligence: machines can learn, but can not understand

The robot will never become human: the sad story of electronics

About artificial intelligence today is not only lazy. But if you give an example Siri, Alice or Alex, for example, the spell check on mobile devices, we are not talking about a full-fledged artificial intelligence. It’s just a highly specialized program.

Computers can’t “think”

Every time one or the other company says that produces a new function based on the AI, this usually means that it uses the principle of machine learning to build neural networks. “Machine learning” is a technology that allows the car to “learn” to more effectively perform a specific task.

We in any case do not criticize the principle of machine learning! It’s a fantastic technology with a large range of applications. But this is not a universal artificial intelligence, and if we look at the obstacles facing today’s machine learning, we can understand why modern technology is the AI so limited in their capabilities.

Artificial intelligence in science fiction is usually a computer or robotic brain that thinks about all kinds of things and understands them in the same way as humans. Such a cyber-intelligence can be attributed to the category of AGI (Artificial General Intelligence), that is, he can talk about different things and use these arguments in many different areas. Typically, this term is replaced by a more simple concept – “strong AI”, which essentially means that the machine can think like a human.

People have not invented such technology. In fact, we are not even close to it did not come. Computer programs like Siri, Cortana or Alexa do not take information and don’t think like we humans. In fact, they don’t “understand”.

Artificial intelligence, which we see now is well trained to perform a specific task with the condition that people will be able to provide a set of data, based on which he will study. This intelligence is learning to do certain things, he still doesn’t understand what he does.

Computers can not understand

Gmail has a new Smart Reply (“smart answers”), which provides ready answers to emails. Smart Reply took the phrase “sent from my iPhone” as a normal response. She also wanted to offer me the phrase “I love you” as a response to various emails, among whom were business emails.

The fact that the computer is not aware of what is written in the letter. Recently it became known that many people send the replies by email. The program does not understand that you do not want to write “I love you” in a response letter to your boss.

Another example: Google Photos did a collage of random pictures of home carpet. Then she spotted this collage in the album “Recent Highlights”, which automatically collects important and meaningful photo, and brought it on the screen of our Google Home Hub. Google Photos to recognize pictures, but didn’t understand what is significant and what is not.

Machine learning cheat

Machine learning allows you to define the problem and give the computer the ability to choose the most effective ways of doing it. Because the machines do not understand what they are doing, ultimately we can “teach” them perfect to solve the wrong problem, which I thought initially.

Here is a list of funny examples where “artificial intelligence” was to play games and perform certain tasks, but instead was looking for in the system of loopholes. All of these examples taken from this interesting table:

• “To develop speed, being very grown up and created her own falling to the ground in the direction of movement”.
• “Agent kills himself at the end of the first level to not die on the second.”
• “Agent puts the game on pause, so as not to lose.”
• “The simulation of life, where survival demanded energy, and the birth of children with no, one kind of developed a sedentary lifestyle that consisted mainly of mating with the purpose of the birth of new children, which then make it possible to eat (or use for production of more edible children)”.
• “Because AI will probably be “killed” in case of loss, the possibility of failure and collapse games was identified as an advantage from the point of view of the process of genetic selection. Therefore, some AI quickly learned to do it.”
• “Neural networks are created to classify edible and poisonous mushrooms, used the fact that the data is alternated, and actually stopped to analyze downloaded image”.

Some of the solutions may seem rather far-sighted, but still none of these neural networks do not really know what I was doing. They set a goal, and they have learned to achieve it. If the goal is to avoid defeat in a computer game, the pause button seems to them the easiest and fastest solution.

Machine learning and neural networks

In machine learning the computer is not programmed to perform a specific task. Instead, it receives the data and evaluates their relevance from the point of view of task execution.

The simplest example of machine learning is a function of image recognition. Let’s say we want to teach a computer program to recognize photos that have a dog. We can provide the computer, millions of images with dogs and without. Image will be labeled – “a dog”/”no dogs”. A computer program “learns” to recognize dogs based on this dataset.

This technology is used to improve neural network, which is a computer program made of several layers, through which all the data sets. Each of these layers assigns according to the different weight and the probability, after which a decision is made. This technique is based on our views on the work of his own brain, where layers of neurons are involved in thinking about the problem. The “deep learning” typically describes a neural network with many layers separating the point of entry and exit data.

Because we know exactly what photos, and what – not, we can skip them through the neural network and verify the correctness of the answer. If the network decides that a particular dog photos there, although she’s actually there, she was prescribed a special algorithm that reports to the network that she was wrong and needed correction before a new attempt. So the computer learns how to better determine if there is a photo of the dog.

It all happens automatically. Having good software and a large amount of structured data that can be used for learning, the computer can adjust its neural network to find an exact match dogs in the photos. This is what we call “AI”.

But, in the end, you still do not get smart program to understand what a dog. This will be the only machine that has learned to determine if the photo dog. These abilities are also impressive, but it is not the ultimate dream.

Besides, based on the data provided, the neural network may be not so “smart” as you would expect. For example, if your dataset has no images of cats, a neural network can fail to see the differences between cats and dogs, and simply celebrates all dogs like cats.

Apply machine learning?

Machine learning is used to solve a variety of problems, including speech recognition. Thanks to these technologies, voice assistants like Google, Alexa and Siri has learned perfectly “understand” what people say. Their algorithm analyses the huge amount of data with samples of human speech, and so the machines all have a better understanding of what sounds make up certain words.

Unmanned vehicles utilize principles of machine learning in order to properly recognize objects on the road and properly respond to them. In the Google Photo there are plenty of features like “live albums” that automatically detects people and animals in photos by using machine learning technology.

DeepMind has used this technology to create AlphaGo – a computer program that learned to play a complex Board game Go and win the strongest players in the world. Machine learning used to create programs that successfully play in other games – from chess to DOTA 2.

Technology is used to identify the person in the latest models of iPhone. Your iPhone creates a neural network that learns to recognize your face. To do this, Apple had to incorporate into their gadgets special chip that performs all the complex calculations related to machine learning.

Machine learning can be used for many other things, from detecting credit card fraud to the preparation of individual recommendations to customers on a shopping site.

However, the neural network functioning on machine learning technology, still do not understand. It’s just a useful program which can perform highly specialized tasks and no more.