Shulou(Shulou.com)11/24 Report--

In 47025.4, Starfleet's interstellar robot lieutenant commander Dita was recruited by his rebellious robot "brother" Lowe to join the rebellious humans-which made the captain of the Enterprise Jean-Luc Picard very uneasy. "the reign of carbon-based creatures is coming to an end," Lowe told Picard. "you, Picard, and people like you are out of date."

This is the optimism expressed in Star Trek that it will take at least three centuries for machines to depose humans. This is just the case in science fiction movies, but in real life, intelligent machines have taken the first step to take over the world, and their time has come.

As a specific sub-field of the broad concept of artificial intelligence, machine learning has penetrated into many human fields, from medical diagnosis to the search for new subatomic particles. Deep learning is the most powerful embodiment of machine learning, which has applications in speech recognition, language translation, image recognition, car driving, new material design, stock market trend prediction and so on.

Computational neuroscientist Searle wrote in an article: "because computers can process vast amounts of data effortlessly, deep learning can not only reshape modern society." it could also revolutionize science-covering subjects ranging from particle physics and organic chemistry to biological research and biomedical applications. "

In recent years, a large number of new papers on machine learning, deep learning and artificial intelligence have flooded into the scientific literature library. A review of this new research covers the fields of health care, epidemiology, material science, basic physics, quantum computing, molecular interaction simulation, fluid mechanics, clinical psychology, economics, visual science and drug discovery.

These reviews highlight the important achievements of machine learning and the prediction of possible follow-up achievements, and most articles also mention the limitations of intelligent machines. For example, some impressive correct answers come from learning shortcuts rather than real understanding.

As a result, intelligent machines are obviously easy to be misled. Moreover, most of today's so-called intelligent machines are confined to a certain field of 1st Skill, dealing with specific tasks efficiently, rather than having the flexibility of general human cognitive ability. For example, a computer can beat a master in a chess game, but can't play poker.

Computer scientist Melanie Mitchell wrote in her book artificial Intelligence: a Guide to Human thinking: "in contrast to humans, most of today's' learning'of artificial intelligence cannot be transformed between related tasks."

Michelle explained that there are still many obstacles to the real exploration of artificial intelligence-machines still cannot think and reason about the world as much as (at least some) humans do.

"people tend to overestimate the advanced nature of artificial intelligence and underestimate the complexity of their intelligence," Michelle said. " The fear of super-smart machines taking over the world is misplaced, she quotes a behavioural economist as saying: "We should indeed be afraid, but not afraid of intelligent machines, but to make decisions about machines in directions that their intelligence cannot understand. I am more afraid of machine mental retardation than machine intelligence."

Machine learning is making rapid progress. In fact, computer scientists have developed some quite powerful algorithms to train machine learning. Usually these learning relies on various computing systems called neural networks. These neural networks are roughly modeled on the processing units of human brain nerve cells. In a traditional neural network, one layer of artificial nerve modifies the correlation strength with another layer of nerve by receiving the input signal, so as to identify and transmit the input signal to the output layer. So the artificial neural network can "learn" to recognize the input signal as a picture of a cat.

In the past decade or so, mainstream machine learning strategies rely on multi-layer artificial neural networks, which is also known as deep learning. A deep learning machine can recognize different patterns in patterns, making the classification of input more refined, even beyond the ability of professionals. A trained deep learning system can even identify cancer signals that have escaped the eyes of radiologists in CT scans.

In some systems, machine learning is "supervised", in which case the training machine uses tagged data, while unsupervised learning means that the training machine is not told the meaning of the large data set entered. The computer itself needs to identify patterns of classification or behavior. In another method called reinforcement learning, if the machine can achieve a goal (like winning a game), it will be "rewarded" when processing input (such as adding points in a storage file). Reinforcement learning shows its power by helping machines beat humans in go games.

Although machine learning deserves a headline in go, the practical achievements in medicine, industry and science are even more remarkable.

2016 South Korean professional go player Lee se-dol faces Google's artificial intelligence program AlphaGo during a go match. In the last game, Huang Shijie, chief programmer of Google Deepmind on the left-hand side, is dropping his first chess piece. Alpha go wins most games against humans as well as defeats. CREDIT: in AP PHOTO / LEE JIN-MAN medicine, machine learning helps researchers improve the weaknesses of standard tests of therapeutic effectiveness. Medical trials that test the effectiveness of disease treatment usually rely on the average of test results to determine effectiveness, so they may miss ingredients that are beneficial to a small number of patients. For example, a treatment found that a weight loss program did not reduce heart problems in diabetics, but machine learning algorithms did identify a small number of patients with reduced heart problems during weight loss, according to epidemiologist Wimken and computer scientist Kelly.

At the same time, machine learning is also helpful in testing the discovery of new drugs. "Deep learning is widely used in drug discovery methods," chemist Zhu Hao wrote in a recent annual review of pharmacology and toxicology. "in this big data era, recent advances in artificial intelligence supported by deep learning show great prospects in the field of drug discovery."

Similar to the discovery of new drugs, machine learning has shown productivity in discovering new materials for industrial applications: machine learning algorithms can simplify the process of finding wear-resistant and tension-resistant "superhard" materials. Material scientist Sparks commented in an article: "this study …"... it is a wonderful manifestation of the important role of machine learning in discovering new structural materials. "

I am more afraid of machine intelligence than machine intelligence, in addition to widely concerned practical applications, machine learning also provides convenience for basic scientific research. In high-energy particle accelerators such as the large Hadron Collider (LHC), proton collisions produce complex streams of particles containing other subatomic particles (such as the famous Higgs boson discovered by LHC in 2012). Scientists need to wisely pick out valuable content from beams that collide with billions of protons per second, which is equivalent to deciding which molecules to inhale when drinking water with a fire high-pressure water gun. Machine learning can help distinguish valuable events from background noise. Other machine algorithms can help identify particles in the impact debris.

Physicist Geist and colleagues said: "Deep learning has affected LHC data analysis and set off a new wave of cooperation between machine learning and particle physics."

Kallio, a quantum physicist, mentioned in another article that the application of machine learning in data processing appears not only in the field of particle physics, but also in cosmology, quantum computing and other basic physics research fields.

"with the rise of machine learning technology in industrial applications, scientists begin to look for the potential of machine learning in basic research." Callio and his collaborators wrote in a review last year.

The limitations of learning, as Kallio and other review authors have stressed, machine learning also has its shortcomings, and scientists should not be blinded by its achievements: "have a healthy and critical attitude towards the potential and limitations of machine learning." including analyzing the breakthroughs of these methods and what they are obviously not good at. "

On the one hand, the "intelligence" of a machine is limited to the nature of its learning data. For example, trained machines can screen job seekers by analyzing human employment decisions and create biases against specific groups of people through historical discrimination data.

Even if the machine performs well, it just looks smarter. For example, reports on image recognition should take into account that the accuracy of the machine usually refers to its first five "guesses"-if any of the five is recognized correctly, the machine will be rewarded.

Sometimes a seemingly intelligent machine performs tasks accurately not because it understands it as humans do, but because it finds a shortcut to the right answer. "A deep neural network that seems to easily identify a cow may not be able to recognize a cow that appears outside the pasture scene." In this case, "pasture" becomes a shortcut for the system to identify "cattle," Gelios and his collaborators wrote in a recent preprinted article.

Sometimes the machine uses texture rather than shape as a shortcut to identify objects, and if a cat's picture is transformed into a relief image with gray shadow through image processing software, it may be recognized as an elephant.

Such shortcuts may be why machines are easily fooled by hostile deception.

Michelle commented in her book: "it is surprisingly easy for humans to secretly deceive deep neural networks." A small change in medical X-rays that the human eye will ignore may change the result of a machine diagnosis from 99% cancer-free confidence to 99% cancer-free confidence.

Michelle believes that because humans do not understand how machines make decisions, it is difficult to explain these mistakes. Operation in a multi-layer deep learning neural network is like operating in a black box, which can not be perceived by human beings, so it is difficult to determine the operating principle of deep learning.

"because decisions made by deep neural networks are often difficult to understand, their mistakes are difficult to predict and fix," she added. "

No matter how the machine learns, it is different from the human way. Not only machine intelligence is difficult to understand, but also human intelligence. Only when scientists have a more complete understanding of human intelligence can they develop a more powerful substitute for artificial intelligence. So we don't have to worry too much that Dita's brother Raul will rule the human race soon.

Michelle shared the view of software entrepreneur Kapoor: "Human beings can only be a wonderful, subtle and rarely explored phenomenon, and there is no danger of replicating it." And this may be the case even in the 24th century.

Author: Tom Siegfried

Translation: zhenni

Revision: Tibetan idiots

Original link: Why some AI is smart until it's dumb

This article comes from the official account of Wechat: Institute of Physics, Chinese Academy of Sciences (ID:cas-iop), author: Tom Siegfried

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.