Driverless cars that automatically avoid obstacles, smartphone voice assistants that "flirt" with when we are bored, recommendation algorithms that we love better than friends and family... I don't know if you have noticed that artificial intelligence technology has long penetrated into all aspects of our lives, and the era we are now in is approaching the next technological revolution at an unprecedented speed, and the key to opening the door of the next technological revolution is hidden in the vast blue ocean of artificial intelligence.
Part.1
Machine learning also has to lay a good foundation first
In the face of automatically turning on and off household appliances, and smart speakers to talk, perhaps you will be surprised by their intimate intelligence, how do machines read our ideas? In fact, our current artificial intelligence is basically based on machine learning technology.
Machine learning, as the name suggests, is to give machines the ability to learn, and learning ability is the watershed between ordinary machines and artificial intelligence. Let's go back to the time of James Watt's improved steam engine in 1765, and no matter the spring, summer, autumn, or winter, the machine that started mankind's first industrial revolution would only repeat an action day and night, and no one who saw the machine would associate it with the word "intelligence." With the advancement of technology, many more complex machines have been designed, but they still do not have basic intelligence.
This continued until the 1950s, when computer theory developed rapidly and machine learning finally took to the stage of history. The basic idea of machine learning is actually not complicated, we train the algorithm to input a large amount of data, let the algorithm generate a model to achieve its potential laws to reveal and predict the future situation. Imagine that starting tomorrow, you try to predict the behavior of the sprinkler passing by in front of your house, and for the first six days, you find that the sprinkler will pass by at five o'clock every day, and then you naturally form a simple understanding (model): the sprinkler will pass by at five o'clock every day. If the seventh day is Sunday, the sprinkler is on holiday (but you don't know), and you find that the sprinkler is not passing by as usual, which means that the previous model is not completely accurate, so after another week, the sprinkler truck is still five o'clock every day for the first six days, and Sunday does not appear again, you can correct the model with new data, so as to get closer to the truth.
This process is the same as our learning process, at the beginning, our machine is like a blank piece of paper, nothing is known, just as before the first day we did not know that the sprinkler truck will come, but it does not matter, when we try to input data to the machine with machine learning ability, everything becomes different, assuming that the protagonist of the above story is a very lazy person, he is not willing to use his brain to think about when the sprinkler will come, so he hopes to let machine learning help him predict the behavior of the sprinkler, He inputs the behavior data of the sprinkler truck into the algorithm every day, a process we call "training", and through a large amount of data training, our machine learning algorithm will become more and more accurate in predicting the behavior of the sprinkler.
Figure 1 The machine learning process is similar and similar to the human learning process
Part.2
"Three years of college entrance examination five-year simulation", the machine can not hide from the sea of tactics
With the data, we also need to choose the appropriate "learning method" to make AI learn faster and better. You may have heard some nouns related to machine learning to a greater or lesser extent, but you are confused, such as supervised learning, reinforcement learning, etc., in fact, these are different training methods that describe the machine learning process, and often apply to different situations.
For example, we want an algorithm to learn to recognize cats and dogs, and if we show the algorithm a large number of photos of cats and dogs in advance and tell it whether the photo is a cat or a dog, then this is called supervised learning; if we give the algorithm a large number of pictures of cats and dogs, but do not tell it which cats are dogs and which are dogs, but let the algorithm automatically look for the difference between cats and dogs, this is called Unsupervised Learning. If the algorithm continues to do multiple choice questions, each time let the algorithm look at the picture to choose whether it is a cat or a dog, the correct answer rewards extra points, the wrong answer penalty deduction points, the algorithm in trying to score as much as possible, avoid the deduction of points after a lot of training will "evolve" their ability to correctly identify cats and dogs, which is reinforcement learning.
So see here you should probably be able to guess how we make machine learning work, it is through a lot of data training, the machine can have such a powerful ability, even if the scientists behind Alpha Go are not Go masters, or even completely do not know Go, but also let Alpha Go win the world championship, and this is impossible to happen in the traditional machine, because all its behavior is written in advance by the designer, so it can not achieve behavior beyond the designer's cognition.
Part.3
Machine learning: Not invulnerable
Seeing this, you may want to ask: "So, does machine learning mean that we are close to the era of artificial intelligence?" ”
Unfortunately, we are still far from being truly strong AI, because current data-based AI algorithms are often very limited. For example, an AI that has been trained to identify cats and dogs for a long time may mistake chihuahuas for cats, and may also mistake hairless cats for dogs, mainly because the results obtained by machine learning are highly correlated with training data, if the data used for training AI is biased, such as the photos of cats used for training are basically hairy long-tailed cats, and the photos of dogs are basically large dogs, then the AI obtained by this training will be easy to make mistakes in identifying some other kinds of cats and dogs.
The more fatal problem with current AI is that many times, due to the interpretability flaw of machine learning, it is a black box process, we can not explain what characteristics it is based on to make judgments, we humans in learning to identify cats and dogs, often focus on certain key parts of the judgment features of cats and dogs, but an AI obtained through image training, even if the results are highly correct, it is possible to put some judgment characteristics on the environment, which is obviously unreasonable. Such features lead to potential risks in the application of AI, such as an autonomous driving AI If we can't tell what it is based on to make driving decisions, then even if it is extremely safe in pre-promotion tests, it may make fatal mistakes in the complex road conditions that appear in reality, and a car accident in the United States last March was because the auxiliary self-driving system mistakenly identified the white car of the truck as the sky, causing the car to crash straight into it.
Interestingly, this also clarifies a truth for us from another perspective: although the sea tactic is useful, it is not efficient, and it will lead to potential errors, and if you want to fundamentally learn new knowledge, you must apply causal logic to fundamentally understand the ins and outs of things, which is what scientists currently hope to achieve in artificial intelligence.
Part.4
To implement causal logic, the machine still needs to work
Judea Pearl, founder of bayesian networks and winner of the Turing Award, believes that the key to making artificial intelligence achieve an essential leap lies in everyone's brain, and the most powerful weapon god has given us human beings - causal logic.
Pearl divides causal thinking into three levels: the first level is correlation, corresponding to the ability to observe, which is the level at which our current data-based weak artificial intelligence is located; the second level is intervention, corresponding to the ability to control variables to carry out actions, that is, the ability to obtain cognition with the help of intervention; the third level is counterfactuality, corresponding to the ability to imagine. Fortunately, all of us are at the third level, and imagination gives us the ability to construct cognition through imagination to construct counterfactual—that is, imaginary worlds," such as the famous elevator thought experiment that led Einstein to generalize special relativity to non-inertial frames with acceleration a century ago.
Figure 2 Pearl's "Causal Ladder"
(Image from Judea Pearl's book Why: A New Science of Causality)
The difference between correlation and causation is that association is the most superficial information between data, that is, correlation. For example, the data shows that the temperature of the year is correlated with the crime rate, and the crime rate is higher when the temperature is low, and if we input these data into AI that only understands the correlation, it is prone to making mistakes when predicting the crime rate. For example, the reason for the increase in crime during the Spring Festival is mainly because the activities of thieves become frequent during the Spring Festival, and the Spring Festival is generally the time when the temperature is the lowest, so if we only analyze the data from the perspective of correlation, we will get the result of the increase in the crime rate caused by the decrease in temperature. If we use this AI to predict the crime rate of a country without a Spring Festival culture, or the crime rate of a year with abnormal temperatures, we will obviously get the wrong conclusion.
But from a causal point of view, we must not only analyze the correlation between the data, but also judge the internal logical chain, for example, when the temperature remains the same throughout the year, will the crime rate change? If the answer is "yes", then we think that there are other influencing factors in addition to temperature, such as last year due to the epidemic, the Spring Festival flow of people decreased, although the winter temperature still fell as usual, but the crime rate did not change with it, then we think that the Spring Festival flow of people is the main reason for the change in the crime rate.
Pearl believes that an important channel for rising from machine learning to causal learning is the introduction of the "intervention" (do) operator, P(A| B) is completely different from P(A|do B). The reason for this is that "intervention" and "observation" are fundamentally different in nature, for example, observing that the rooster is chirping and forcing the rooster to crow are two completely different things, and our current algorithm can easily answer the correlation between the two things of the rooster chirping and the sun rising, but it is difficult to correctly answer whether the sun will also rise when the rooster is forced to crow. Pearl argues that artificial intelligence that only accepts passive observation data cannot climb the second ladder to answer intervention-related questions, and therefore cannot understand the causal relationship between "rooster chirping" and "sun rising", because the confirmation of causality requires control variable experiments, and such experiments themselves are based on intervention. You may ask, if the observation dimensions are large enough, is it a substitute for intervention to obtain sufficient data? In fact, it is difficult to ensure that the data range is consistent with the actual test environment, and the more difficult part is that many times it is impossible to know whether the data itself is complete or not, which leads to the algorithm that we train with a huge amount of data, which may be wrong because the data is not completely consistent with the test environment, which is called the OOD (Out of Distribution) problem, even turing award winner Yoshua Bengio also believes that OOD generalization is the most urgent problem for current artificial intelligence to solve.
Seeing this, you can probably understand why we are still far from imaginary artificial intelligence, because the artificial intelligence we are currently building, let alone has a third-level imagination, does not even have the ability to judge cause and effect. Fortunately, scientists have realized that causal learning is the key to making artificial intelligence achieve the next leap, and many scientists have successively invested in theoretical research on causal machines, such as Professor Cui Peng of Tsinghua University proposed stable learning that combines causal reasoning with machine learning to improve OOD generalization problems, and Dr. Huang Biwei of CMU used causal discoveries to achieve more accurate predictions on the unsteady data of time series. I believe that with the development of technology, artificial intelligence technology will become more and more reliable and benefit mankind!
Source: Voice of the Chinese Academy of Sciences
This article has been reprinted with permission, if you need to reprint, please contact the original author
The article only represents the views of the author and does not represent the position of the China Science Expo
Reprint indicating source Unauthorized reprinting shall not be reproduced
China Science Popularization Expo is a science popularization cloud platform of the Chinese Academy of Sciences, sponsored by the Computer Network Information Center of the Chinese Academy of Sciences, relying on the high-end scientific resources of the Chinese Academy of Sciences, committed to disseminating cutting-edge scientific knowledge and providing interesting scientific and educational services.
Click here to tell me you're watching