Is mankind in danger?

But there was something completely different about Google models, Tesla and General Motors, demonstrating the increased power of artificial intelligence. Navia's car did not follow the instructions of an engineer or a programmer, but fully trusted a algorithm that had learned to drive by observing a human being. The design of a machine of this kind is an extraordinary but disturbing enterprise, as it is not entirely clear how the machine makes its decisions.

Information recorded by sensors on a large network of artificial neurons that process data and provide instructions needed to use the wheel, brakes, and other systems. The results look similar to the behavior expected by a driver in the flesh and bones. But what would happen if one day the car made an unexpected move, such as going and crashing into a tree or stopping with the green light? In the current situation, it would be very difficult to understand why it happened. The system is so complicated that even designer engineers have found it hard to individualize motives in the origin of the decision. It is also impossible to ask for an explanation of cars: there is no specific criterion to design the system so that it can explain why it does what it does.

I thought the mysterious thought of this car leads to an open question about artificial intelligence. The underlying technology, known as the deep lesson, has in recent years been very effective in solving problems and has often been used for purposes like translation, vocal recognition, and image selection. It is now expected that the same techniques will be able to diagnose deadly diseases, make millions of investments, and revolutionise entire industrial sectors. But all of this would not happen (or at least shouldn't happen) if there wasn't going to be a way to make techniques like dep teaching more understandable to their creators and the primary responsibilities for what they use.

Otherwise it will be difficult to predict when there will be problems when they will be inevitable. Not by chance, Navia's car is still in experimental phase. Today, mathematical models are already used in the United States to select, for example, who can obtain the controlled freedom, who can borrow a loan, and who should be employed for a job. By entering these models, you can understand their way of reasoning.

Recently, however, banks, armies, businesses and other subjects pay attention to new complex automatic-understanding systems that are at risk of making completely impervious automable decision-making processes. Deep Learning, the most popular system, is a completely new way of programming computers. “Qych is now an important problem and will be much more in the future”, notes Tommi Jaakkola, a professor from Massachusetts Institute of Technology (MIT) working for automatic understanding applications.

For any decision in the financial, medical, or military field, you cannot believe just a black box”, that is, a system to understand events only with finished things. One already points out that the possibility of seeking artificial intelligence as achieved in perceived conclusions should be guaranteed by law.

By the summer of 2018, the European Union could impose upon companies the obligation to explain to customers decisions made by automated systems. Perhaps it will result in an impossible enterprise, even for systems that at first look relatively simple, such as apps and websites that use deep warning for advertising or to recommend playlists with songs. Computers managing these services are programmed themselves and no one is able to understand how they do it. Not even the engineers who developed the apps.

All this opens up a series of difficult questions. With technology advanced sooner or later, a threshold beyond which the use of artificial intelligence will require an act of confidence. It is also true that we human beings are sometimes unable to explain our mental problems, but sometimes we trust our intuition to evaluate people and decide whether to believe them or not.

Will they be able to do it with machines that think and make decisions differently from human beings? Never before were cars built capable of doing unintelligible things even from those who created them. How do we plan to communicate and agree with intelligent machines that can result in unpredictable and inadequacy? These questions have prompted me to make a trip to the most advanced research on artificial intelligence, from Google to Apple, including a meeting with one of the great philosophers of our age.

Clinical Review

Nel 2015, a research group of Mount Sinai Hospital in New York, has decided to apply deep lecture to the hospital database, which includes hundreds of information on patients, from clinical examination results to medical visits. There was a program, called Deep Patent, that was designed to use the data of some 7,000 persoses. Where it's tested on new patients, it's been extremely effective in predicting pathology. Without any guidance from the specialists, Deep Patent has discovered repeated elements inside hospital records that were possible to predict when a person was more exposed to a series of diseases, among which tumors in the Montenegrin liver.

There are excellent methods of predicting diseases from the patient's clinical card, says Joel Dudley, head of the Mount Sinai Hospital research group. But that, he adds, “works much better”. In many ways, however, Deep Patent is a mystery. For example, he is able to predict surprisingly well the display of psychiatric concerns such as the work of the diagram. As soon as the outline is hard to predict, Dudley asked how it was possible. He still hasn't found an answer. Deep Patent offers no related indicators. To provide real assistance to doctors, an instrument must provide a rational explanation of its prediction, provide us with accurate reassurance, and perhaps justify the use of various drugs from those given up to that point. We know how to build these models, but we don't know how to operate”, says sad Dudley.

Artificial intelligence hasn't always worked this way. From the beginning, there have been two schools of thought about how understandable or explanatory it should be. For many, the most meaningful thing was to build cars to reason on a series of rules and logic, making their operation transparent to anyone who would like to consider the code. Others emphasized that in fact, intelligence would develop more easily if they had followed the example of biology, learning from observation and experience.

This meant completely overturning computer programming. It was not with the programmer writing the commands for solving a problem, but it was the program that specified the algorithm itself on the basis of the desired examples and outcome. Automatic understanding techniques that evolved into the very powerful systems of modern artificial intelligence have followed the second course: In essence, the machine is programmed itself. At first, this method had limited practical application, and in the 1960 ' s and Seventy ' s, margins remained.

Then, the computerisation of many industrial sectors and the release of large series of data have renewed interest. All of this has led to the development of more evolved automatic techniques and especially in the evolution of technology known as the artificial network of neurons. Since the Ninth ' s, networks were able to understand the characters written by hand.

But only in the beginning of the last decade, after various adjustments and refinings, the large networks of neurons (or <x0) the deep networks of neurons” have provided evidence of substantial improvements in automatic perception. The merit of the current explosion of artificial intelligence is deep teaching, which has given computers extraordinary capacities: for example, the knowledge of language spoken almost as a person in the flesh and bones, a highly complex ability to code it by car.

Deep warning has transformed the artificial vision and substantially improved the informal translation. Today, it is used for important decisions of any kind in medicine, finance, manufacturing, and other sectors. The function of automatic understanding technologies is essentially more opaque than systems based on lines written by a programmer. This does not mean that all future artificial intelligence technologies will be equally unintelligible. But, by its nature, deep warning is an extremely dark “kkt”.

It's not enough to look inside a network of neurons to figure out how it works. A network's reasoning scheme is embedded in the behavior of thousands of simulated neurons, organised in dozens or up to hundreds of layered layers. Every neuron in the first layer gets an input for example the intensity of a pixels of an image and makes an account before emitting new signals. In turn, these have been transferred through a complex network of neurons in another layer and so on until they are not reached with a complex result.

All of this adds to a process known as back-propagate that affects specific neurons' calculations in order to enable the network to learn how a certain outcome is provided. The many sections that make up a deep network allow the network itself to know information at different levels of abstraction. For example, a system designed to recognize dogs, the inferior layer knows basic information such as jacks and colors, the superior layers know more complex features such as hair and eyes, and the highest layer knows the whole of information - the dog. The same system, by simplifying it, applies to other inputs that make the car learn from itself - the sounds that form the words in question, the letters, and the lips that make up the inside of a key or the movement of the steering wheel needed to drive.

Very sophisticated systems have been used to find out in detail what happens in these systems. In 2015, Google researchers modified an algorithm for recognition of images based on deep teaching so that instead of recognising objects in photos, it would be able to generate or modify. In fact, by applying the reverse algorithm, researchers have been able to detect information that the program used, for example, for an acquaintance with a bird or a home. The resulting images, produced by a system called Deep Dream, show gratastic animals, similar to aliens, coming out of clouds or trees and hallucinant pagoda that emerge from forests or mountain ranges. Images show that deep teaching is not always entirely unintelligible and reveal that, to recognize birds, algorithms automatically target visible features like beaks and rifles.

But the images make you realize how different it is how different the lesson is from the perception of the human being, starting with the fact that it extracts elements from information that we try to ignore completely. Google researchers have noted, for example, that when algorithm generates the image of a weight lift, it also creates that of an arm that holds it. The conclusion of the car is that the arm is all one piece of shock. Other advances can come from neuroscension and cognitive science. A research team led by Jef Clune, assistant at Wyoming University, has used the equivalent of optical illusions in the field of artificial intelligence to test deep neurons.

2015 Clive has shown that certain images can push the network that perceives things that don't exist, as images use recognition schemes at the minimum level that the system requires. Jason Yosinski, an associate of the Clive, has created an instrument that functions like a probe in the brain. The instrument targets a neuron in the middle of the network and requires an image that activates the most. are all abstract images (try to imagine an impressive representation of a flamingo or a school bag) and reveal the mysterious nature of the machine's perceptive powers.
A Mountain of Functions

These are indications of how artificial intelligence works. You have to learn more, but it's not easy. It's the interaction of calculations within a deep network of neurons that determines more complex knowledge schemes and decision-making processes, but those accounts are a maze of functions and mathematical variables. If the grid were too small, we could decode”, Jaakkola says. “But when it becomes too large, with thousands of units per layer and thousands of layers, it becomes virtually unintelligible”.

In the office next to that of Jaakkola works Regina Barzilay, an MIT Docente who has decided to apply the automatic meaning of medicine. In 2015, at 43 years of age, a breast tumor was diagnosed. The diagnosis itself was traumatic, but it has been even more shocking to him to discover that statistical methods and automatic meaning were not used in oncological research or in the selection of therapies. Barzillai points out that artificial intelligence can revolutionize medicine, but that to take advantage of potentials must go beyond simple clinical cards. Its idea is to use more gross data that is now used a little bit:

“

After the therapy ended, Barzillai and her students have started working with Massachusetts doctors General Hospital for developing a system capable of producing data from patients' notes to individualise those with interesting clinical characteristics for researchers. But Barzillai has realized that the system must be able to explain its reasonings. So, with the help of Jaackola and a student, he added a passage - the system emits and records representative passages of a certain feature. In addition, Barzillai and her students are developing an algorithm deep teaching capable of revealing the first signs of breast tumor in mammographic images and aiming to give this system some justifiable capacity of his reasoning. Must find a vicious circle in which the car and man co-operate”, Barzillai says.

The United States Armed Forces are investing billions in projects that envision the use of vehicles to pilot vehicles and flying vehicles, identify targets, and help analysts filter out large quantities of data. Here more than in any other field, including medicine, there is little room for algorithmic mysteries, and the Department of Defense is busy with the “the basic <x1-x2> transparency. David Gunning, a director of Defence Advanced Research Projects Agency (DARPA), an American government agency that invests in security technologies, is coordinating an extremely appropriate program: Explainable Artificial Intelligence. Gunning explains automation is taking place in many sectors of the armed forces. Secret services analysts are testing the automatic sense to individualise schemes in the large amount of surveillance data in their possession. Many land and air vehicles are in development and testing phase.

But soldiers, probably, would not feel comfortable in robotized tanks that are unable to explain their decisions, and analysts would hardly use information in the absence of a scheme of reasoning. For this reason, these automatic understanding systems often create false alarms, so analysts need some more element to understand why it is such an indicator”, Gunning says.
In March DARPA has selected 13 academic and business projects to finance under Gunning's programme. Some of these projects can be used by Carlos Guestrin, a professor at the Washington University. The Guestrin Group has developed a system that enables automatic understanding algorithms to justify their results. Basically, the computer automatically extracts from the examples of a data series and uses them to provide a quick explanation. For example, a system designed to classify e - mails from terrorists normally uses different emails in the process and understanding and decision. And the Washington University system is able to isolate some key words in a message. Guestrin and his colleagues have also found a system that enables algorithms to know images to understand their reasoning scheme by recording more significant pieces of an image. The problem of this and other similar systems, for example, that of Barzillai is that explanations have always been simplified. This means that some important information is lost.

Your dream hasn't been realized yet. The objective is to build an artificial intelligence that is capable of interacting with human being and explaining its behavior”, Guestrin says. “We're still away from an artificial intelligence that is truly ineffable”. I knew how artificial intelligence argues would be fundamental if technology really became an integral part of our daily lives. Tom Gruber, head of the team that is developing Apple's virtual Syrian assistant, says commentary is a key objective of his group.

He and his colleagues are working to make Syria more intelligent and capable. Gruber does not talk about future developments, but it is easy to imagine that if Syria advises us a restaurant, we would like to know why.

Ruslan Salakhdinov, director of Research on Artificial Intelligence at Apple and Associated Professor at Carnegie Mellon University of Pittsburgh in the United States, considers the heart of the report between human beings and intelligent vehicles. As with human behavior, artificial intelligence may not be possible to simplify everything it does. Even if someone gives us an explanation that seems reasonable, it allows it to be incomplete and the same can serve artificial intelligence”, says Clive. For its nature, intelligence is probably rationally only partially explained.

There's a part that's just instinctive, involuntarily, non-student”. If things are like this, at a certain moment we will either have to trust artificial intelligence, or we will have to give up using it. Artificial intelligence will become part social intelligence. The social contract is based on a series of expected positions and therefore artificial intelligence systems will have to be designed to adapt to our social norms. If we really build robot tanks and other cars capable of killing, it's essential that their decision-making processes be consistent with our ethical judgments.

Encycloperic Traffices

I went to the Boston Tufts University, where I met philosopher Daniel Dentt. A chapter of his last book “From bacteria to Bach and back”, a type of encyclopedia conscience treaty, speculates that some of the evolution of intelligence will be the creation of systems capable of developing activities that anyone who created them cannot do. The question is: What measures should we take to do things calmly, what standards should we ask for these systems and ourselves?”, the philosopher asks.

Dennett thus warns of the risks associated with the research of extenuation. If we have to use these machines and have to give them maximum control over how and why they respond to us, he says. But since there's probably no perfect answers, we need to avoid trusting artificial intelligence explanations the way we don't trust those of our likes, no matter how intelligent the car would be. “If artificial intelligence is unable to explain better than what it does”, it concludes, “then we shouldn't believe”.
(Will Knight for MIT Technology Review)