Recently, UC Berkeley University School of Electrical Engineering and Computing and the School of Statistics, a master of machine learning, Michael I. Jordan, known as "Michael Jordan of the artificial intelligence community," published an article reminding everyone that not only deep learning is not All of "artificial intelligence", even the "artificial intelligence" we discuss every day, is not the whole "artificial intelligence" in the true sense; there are not only many problems that we have neglected, but also for the benefit of all mankind. The new discipline of how to build artificial intelligence systems has just sprouted and is waiting to be established.
Artificial intelligence (AI) is the mantra of people of this era, and it has been repeatedly chanted by technical experts, scholars, journalists and investors. Just as there have been many phrases in the past that have spread from the academic field of technology to the general public, there is a serious misunderstanding when people use the phrase AI. It used to be because the public didn't understand the scientists, but this time scientists are as confused as the public. The possibility of silicon-based wisdom that has the same intelligence as ours in this era makes us all interesting, it attracts us and scares us. However, it will also distract us.
For what happened in this era, I want to tell a story with a unique perspective. This story involves humans, computers, data, and life-and-death decisions, but the focus here is on something other than the fantasy of silicon-based intelligence. 14 years ago, when my wife was pregnant, we did an ultrasound. The geneticist who examined her pointed out that there were some white spots around the fetal heart. “These are signs of Down's syndrome,†she said. “The risk of illness has risen to one-twentieth.†She further told us that amniocentesis can be used to check whether the fetus has a Down syndrome mutation. gene. However, there is a risk of amniocentesis, and the mortality rate of the fetus during surgery is about one-third.
As a statistician, I decided to find out the source of these numbers. To make a long story short, I found a statistical analysis of the disease 10 years ago from the UK, which believed that these white spots reflect calcium accumulation and are one of the predictors of Down's syndrome. But I also noticed that my wife's imaging machine was a few hundred pixels more per square inch than the machine used in the UK study. I went back and told the geneticist that these white spots were probably false positives - they were actually "white noise." She said, "Oh, then I know why the Down syndrome that we diagnosed suddenly changed a few years ago. We were the new machine that was changed at that time."
We did not end with amniocentesis. A few months later, our daughter was born and very healthy. But this incident made my heart confusing, especially after a rough calculation. I am sure that on the day the doctor told us the diagnosis, thousands of people around the world got the same diagnosis as us. Many of them have chosen amniocentesis and many babies die unnecessarily. This kind of thing happens every day until someone finds out why. The medical problem reflected in this incident is not the only one I will encounter. This is the problem of the entire medical system - measuring variables, obtaining results, performing statistical analysis at some time and place, and then using these in other time and place. in conclusion.
To be precise, the problem lies not only in the data analysis itself, but also in the “provenance†that database researchers call – in a broad sense, where the data appears, what inferences are derived from the data, and the inferences and currents. How big is the situation? Although a professionally trained person may be able to analyze and solve specific situations for each situation, the real problem to be solved is how medical systems used worldwide can do this without the need for sophisticated human supervision. .
I am also a computer scientist, and I suddenly realized that the discipline of establishing this world-scale reasoning and decision-making system—incorporating computer science with statistics and taking into account human experience—has never been taught in any school. . I also realized that not only in the medical field, but also in the fields of business, transportation, and education, the establishment of this rule is at least as important as building an AI system (a dazzling game AI and motion perception system).
Whether we can understand "intelligence" in a short period of time, we all face a huge challenge, that is, how to combine computers and humans to make humans survive better. This problem is seen by some as the birth of “artificial intelligenceâ€, but we can also regard it as a new branch of engineering discipline with a sense of peace and awe.
Just like civil engineering and chemical engineering decades ago, the goal of this new discipline is to bring together the power of some key ideas to safely bring new resources and capabilities to people. Just as civil engineering and chemical engineering are based on physics theory and chemistry theory, this new discipline is built on the ideas we discovered in the last century, such as "information", "algorithm", "data" On the concepts of “uncertaintyâ€, “calculationâ€, “reasoning†and “optimizationâ€. In addition, since most of the focus of this discipline comes from humans and human data, its development also requires the help of social sciences and humanities.
Although some basic theories of this discipline have gradually emerged, the laws that bring them together are still missing. These basic theories can only be linked and piled together individually.
Just as people have been building houses and building bridges before the emergence of civil engineering disciplines, people are now building a social scale that includes machines, people, and the environment without establishing new disciplines. Reasoning and decision making systems. Similarly, just as early buildings and bridges sometimes collapsed in completely unpredictable ways, with tragic consequences, many of our early social-scale reasoning and decision-making systems have exposed serious conceptual problems.
Even more embarrassing is that we humans are not very good at predicting where serious problems will occur next time. We now lack its corresponding engineering discipline and lack the rules of analysis and design.
When the public talks about these issues, the term "artificial intelligence" is used to broadly cover all the concepts related to wisdom, which makes the scope and consequences of emerging science and technology difficult to discuss. Let's take a closer look at what the term "artificial intelligence" means in the near future and in history.
What today's "artificial intelligence" refers in most cases, especially in public discussions, is what we call "machine learning Machine Learning" in the past few decades. Machine learning is a discipline of research algorithms that draws ideas from statistics, computer science, and other disciplines to design algorithms that can process data, make predictions, and help human decision making. As for the impact on the real world, the impact of machine learning is real, and it is far more than just a recent impact. In fact, as early as the early 1990s, machine learning showed obvious signs that could have a huge impact on the industry. In the 21st century, forward-looking companies like Amazon have used machine learning. In the ups and downs of the company's business, it deals with back-end issues such as fraud detection and logical chain prediction, and also builds user-oriented innovative services such as recommendation systems. As the size of data sets and computing resources have grown by leaps over the past 20 years, we can now clearly see that not just Amazon, almost any company that can make decisions based on large-scale data will soon have machine learning as a driving force. . New business models will emerge. The phrase “data science†has also been used to refer to this phenomenon, which reflects the need for machine learning algorithm experts and databases, distributed systems experts to work together to build scalable, robust machine learning systems. The impact of such systems on a larger social and environmental range.
In the past few years, this fusion of ideas and technology trends has also been called "artificial intelligence." However, this title is worthy of our careful examination.
Historically, in the late 1950s, people had a passion for re-emerging human-level wisdom with both hardware and software, and also created the term "artificial intelligence" / "AI." This kind of ambition can be called human-imitaTIve AI. In this concept, an entity with artificial intelligence should be regarded as our partner, even if it does not look like it, it should be mentally. This can be largely seen as an ambition in the field of academic research. Some related academic fields existed at the time, such as operations research, statistics, pattern recognition, informatics, and control science, and they were often inspired by human intelligence (and animal intelligence), but to some extent These disciplines are concerned with "low-level" signals and decisions. For example, the ability of a squirrel to understand the three-dimensional structure of the forest in which it lives, and the ability to jump between branches, are instructive in these disciplines. "Artificial Intelligence" should focus on other things, the ability of "high-level" and "cognitive" in human reasoning and thinking. After sixty years have passed, high-level reasoning and thinking skills are still unpredictable. The technological advances now known as artificial intelligence are largely derived from low-level pattern recognition, motion-related engineering, and the search for patterns in data to make predictions, validate guesses, and make statistical decisions.
In fact, David Rumelhart's back-propagation algorithm, which was rediscovered in the 1980s and is now regarded as the core of the so-called "artificial intelligence revolution", first appeared in the field of control in the 1950s and 1960s. One of its earliest applications at the time was to calculate the thrust of the Apollo spacecraft as it flew to the moon.
Since the 1960s, there have been many breakthroughs in our technology, but to a large extent these advances have not come from the pursuit of artificial human intelligence. Rather, as in the case of the Apollo spacecraft, these ideas are hidden behind the scenes and are the result of research by researchers who are trying to solve some very specific engineering challenges. Although not widely available to the general public, research and system building in document indexing, text categorization, corruption monitoring, recommendation systems, personalized search, social network analysis, planning, diagnostics, and A/B testing are all very successful; Google, The driving force of big companies like Netflix, Facebook, and Amazon is just such technological advancement.
Now we will simply refer to all of these things as "artificial intelligence", and it seems that it is indeed the case. For researchers in the field of optimization or statistics, this kind of classification is a big surprise, and they suddenly become "artificial intelligence researchers." But in addition to the researcher's categorization problem, the bigger problem is that this single, inaccurate abbreviated vocabulary will prevent us from clearly understanding the current large-scale intelligence and commercialization issues.
In the past 20 years, we have made many major breakthroughs. The industry and academia have also created a new kind of thinking, supplemented by artificial human artificial intelligence; we often call it "Intelligence AugmentaTIon". Here, we use computing power and data to build services that enhance human intelligence and creativity. Search engines can be seen as an example of intelligent enhancement. It enhances human memory and enhances human understanding of objective facts. The same is true for natural language translation, which enhances human communication. Computer-based sound and image generation can also be an enhancement of the artist's palette and innovative ideas. However, while such services ultimately inevitably involve high-level reasoning and thinking skills, they are currently a blank in this respect: all they do is through a variety of string matching and numerical values. The calculation finds patterns that humans can use.
Here I also need to come up with a concept that broadly recognizes the discipline of "Intelligent Infrastructure". It refers to a network of computing power, data, and related physical entities that make the human environment more human, interesting, and safer. Such infrastructure has emerged in areas such as logistics, medicine, commerce and finance, affecting countless personal and social activities. Sometimes people talk about the Internet of Things (IoT) will also mention the establishment of some kind of network, but the Internet of Things field only connects the "things" to the "net", how to make these "objects" process the data stream, Discovering information about the world, interacting with humans, and so on, the high-level abstraction of data beyond 0 and 1 is completely untouched.
For example, I can talk about one's own thoughts. We may have imagined living in a “system-wide medical system†that will set up the flow of data and data analysis between the doctor and the medical devices around the patient, thus in the disease. Diagnostics and medical care help human intelligence. This system collects information from body cells, information in DNA, information in blood diagnostics, the environment, population genetics, and information in the vast literature on drugs and medical methods, and then integrates them. It is not about individual patients and doctors, but the relationship between all human beings, just as modern medical experiments are done on a certain part of a person (or an animal) and then healing other people based on the results of the experiment. Similarly, just as modern banking systems can focus on relevance, traceability and reliability in the financial and payment sectors, this medical system is also best to focus on these concepts. And, although we can foresee that building such a system will encounter a variety of problems, including privacy issues, liability issues, security issues, etc., these issues should be correctly viewed as awaiting resolution, rather than blocking The reason for building such a system.
We are now experiencing such a key question: In the face of these greater challenges, is researching human-like artificial intelligence the best way to deal with them, or even the only one? Many of the most frequently mentioned success stories in machine learning are related to human-like artificial intelligence, such as in computer vision, speech recognition, game AI, and robotics. So it seems that we just have to wait for new developments in such areas.
Here I want to point out two things. First, although this is not the case in the newspaper, the research on the direction of human artificial intelligence is actually very limited. We are still far from achieving the goal of real human artificial intelligence. Unfortunately, the limited advances in human-like artificial intelligence are also prone to excitement (and fear), which makes research in this direction too hot and too much media attention. This phenomenon is not seen in any other engineering field. Second, and more importantly, if it is to solve important wisdom enhancement and smart infrastructure issues, the success of human-like artificial intelligence-related fields is neither sufficient nor necessary.
For the sufficiency side, think about driving a car automatically. In order to achieve such a technology, the series of engineering problems awaiting resolution has little to do with the extent to which humans are competent to drive (and the extent to which humans are incapable of driving). An overall transportation system (a smart infrastructure) would be very close to modern air traffic control systems, unlike the current collection of almost ungrouped, forward-looking, careless human drivers. It will be much more complex than current air traffic control systems, especially as it can make use of massive data and adaptive modeling capabilities to make fine-grained decisions. It is this kind of problem that we need to consider first, and for such problems, the efforts in human artificial intelligence will distract us.
For the necessity, some people have proposed that the desire to imitate human artificial intelligence actually includes wisdom enhancement and smart infrastructure, because human-like artificial intelligence will not only solve various classic AI problems (literally, such as Turing test). And it is most likely to address both smart enhancement and smart infrastructure issues. This view can not find any historical precedent as a support. The development of civil engineering depends on how to design artificial painters and masons? Is the discipline framework of chemical engineering creating an artificial chemist? Even more interesting is that if our goal is to build a chemical factory, should we first create an artificial chemist and then let it think about how to build a chemical factory?
There is also a related point of view that human intelligence is the only intelligence we know. The first step in developing artificial intelligence is to try to imitate this intelligence. But there are actually some types of reasoning that humans are not very good at it. Humans have many mistakes, prejudices and restrictions. More importantly, the purpose of human evolution is not to deal with large-scale decision-making problems like modern smart infrastructure, nor is it to deal with uncertainties in a smart infrastructure environment. Some people may say that an artificial intelligence system will not only imitate the wisdom of human beings, but also "correct" it, and of course it can be extended to any large-scale problem. But such an idea went to the realm of science fiction. Such a purely conjective view is of course in line with scientific fantasies, but it should not be the main strategy when we face the emerging problems of intelligent reinforcement and smart infrastructure. We should deal with intelligent reinforcement and smart infrastructure issues in the way they should, rather than just as inferences that mimic human artificial intelligence goals.
It is not difficult to see that the algorithms and infrastructure challenges in smart infrastructure systems are not the core themes in human artificial intelligence research. Smarter infrastructure requires the ability to manage rapidly changing, and potentially unrelated, distributed knowledge stores. Such a system requires interaction between cloud computing and edge computing to make immediate, distributed decisions; it also needs to be able to handle the long tail in the data, that is, there are a lot of data about some individuals, but most Individuals have very little data. They need to be able to handle the problem of sharing data across administrative and competitive boundaries. Finally, and very importantly, smart infrastructure systems need to incorporate economic concepts such as motivation and pricing into the statistical and computing infrastructure that connects people, people and goods. Such a smart infrastructure system is not just about providing services, but more importantly, it provides the market. Some areas such as music, literature, and news are in great need of such a market, where data analysis can connect creators and consumers. And all of this needs to evolve under the premise of social, ethical, and legal.
Of course, the classic human-like artificial intelligence problem is still an important research topic. However, current research in artificial intelligence is based on the collection of data, the deployment of deep learning infrastructure, the ability of these systems to mimic some very specialized human technology, and the inability to explain the rules. The status quo actually distracts us from our attention, leaving us to ignore many of the open problems in classical artificial intelligence. These questions include how to add meaning and reasoning to systems that can handle natural language, how to reason and represent causality, how to develop a representation of computable uncertainty, and how to develop a form that can formalize and pursue long-term goals. System, and so on. These are also classic goals in human artificial intelligence, but in the current "artificial intelligence revolutionary boom", it is easy to forget that these problems have not been resolved.
Intelligent enhancement is still crucial enough. In the case of abstract reasoning of real-world conditions, it is impossible for computers to reach human levels in the foreseeable future. There is a need to build a deep enough communication between humans and computers to solve our most pressing problems. And we also want to use computers to push human creativity to new heights instead of using computers to replace human creativity (in various senses).
John McCarthy (John McCarthy) proposed the term "artificial intelligence" at Dartmouth College, apparently to distinguish him from Norbert Wiener's different research goals. Wiener's word is "cyberneTIcs" to show that his know-how about intelligent systems is closely related to operations research, statistics, pattern recognition, information theory, and control theory. McCarthy is more concerned with the connection between intelligence and logic. But then there was an interesting reversal, and it was Wiener's intellectual goal that ruled the field, but it was the name of McCarthy. (The current status is of course temporary; the wind direction in the AI ​​field is much faster than in other areas)
But for us, the historical perspective of both McCarthy and Wiener needs to be surpassed.
We need to understand that the artificial intelligence that is now being discussed by the general public and focusing on a small part of the problems in industry and academia has great risks that will stop us from paying attention to the full range of artificial intelligence, intelligent enhancement and smart infrastructure. Challenges and issues within.
This scope is not only about some sci-fi dreams and fears about superhuman computers, but more about the need for human beings to understand as science and technology become more and more influential in human life. And control it. And, in this understanding and control, all human beings should make their own voices, not just those who understand technology. Narrowly focusing on human-like artificial intelligence can make many sounds that should be heard unreadable.
Although industry companies continue to bring more technological advancements, the academic community must play a key role. It's not just about providing some innovative technical ideas. It should connect researchers in computing, statistics, and other disciplines that are worth listening to, especially in social sciences, cognitive sciences, and humanities.
On the other hand, although the humanities and natural science disciplines are very important to our progress, we should also bear in mind that our discussion is about an unprecedented scale and scope of engineering projects, and this society needs to build some New artifacts. These artifacts should be constructed to meet the promise. We don't want to find out that some medical, transportation, and business systems have been built before they are actually not good enough to find that they actually reduce human life and happiness. Out of this philosophy, as I have just emphasized, we need a new engineering discipline to guide this data-oriented, learning-oriented field. Although the idea sounds very good, we can't really think of it as a discipline at the moment.
Further, we are witnessing the emergence of new engineering areas, and we should be happy about it. The word "engineering" is often narrowly understood by people. Whether in academia or in more contexts, it seems to be a metaphor for a cold and ruthless machine, or a loss of human control. But in fact, an engineering discipline can be what we want it to be.
In this era, we can really imagine something that has never happened in history: a new human-centered engineering discipline.
I can't give a name to this new subject that is sprouting now, but if the word "artificial intelligence" / "AI" is used as the root of the name of the subject, we must keep in mind the extremely limited reality of the root. significance. Let us broaden our horizons, gather up fanaticism, and carefully observe the challenges we are waiting for.
Ningbo Autrends International Trade Co.,Ltd. , https://www.supermosvape.com