Self-driving cars that can automatically avoid obstacles, smart phone voice assistants that accompany the owner when they are bored, "understand" our recommendation algorithm better than friends and relatives... I don't know if you have noticed, artificial intelligence technology It has already penetrated into all aspects of life, and 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 technology.
Machine learning needs to lay a solid foundation first
When faced with self-operating household appliances or talking with smart speakers, you may feel incredible about their caring and intelligence. How can machines read human thoughts? In fact, the current artificial intelligence is basically based on machine learning technology. It can be said that it is "machine learning" that allows the appliances at home to understand you.
Machine learning, as the name suggests, is to give machines the ability to learn, and the ability to learn is the watershed between ordinary machines and artificial intelligence.
In the middle of the 18th century, when Watt improved the steam engine, the steam engine that started the first industrial revolution of mankind would only repeat one action day and night. Anyone who has seen this machine will not associate it with the word "intelligence". . With the advancement of technology, people have designed many more complex machines, but they still do not have basic intelligence.
In the 1950s, with the rapid development of computer theory, machine learning finally entered the stage of history. The basic idea of machine learning is actually not complicated. It is to input a large amount of data into an artificial intelligence algorithm (AI algorithm), train it, and let the AI algorithm generate a model, so as to reveal the underlying laws of things and predict the future situation.
For example, you want to observe a sprinkler passing by your door and make predictions about its "behavior". In the first 6 days of observation, you found that the sprinkler passed by the door on time at 5:00 every morning. At this time, a simple cognitive model was formed: the sprinkler passed by at 5:00 every morning. If the 7th day is Sunday, which is the day off for the sprinkler driver (but you don't know), and you find that the sprinkler is not passing by as usual, it means that the previous model is not accurate. After another week, the sprinkler is still passing by at 5 a.m. every day from Monday to Saturday, and no longer appears on Sunday. You can use the new data to correct the model and get closer to the truth.
Similarities and differences between machine learning process and human learning process
The process of machine learning is similar to observing a sprinkler. Before the data is entered, the machine is like a blank sheet of paper, knowing nothing, just as you didn't know the sprinkler was coming until the first day. That all changed when researchers tried to feed data to a learning machine. Assuming that the observer of the sprinkler is a very lazy person who is unwilling to use his brain to speculate on the operation law of the sprinkler, he can input the "behavior" data of the sprinkler every day into the AI algorithm of the machine. This process is called "train". Through a large amount of data training, the AI algorithm will become more and more accurate in predicting the "behavior" of the sprinkler.
Machines can't escape the sea of tactics
With the data, researchers also need to choose the appropriate "learning method" to make the machine learn faster and better. You may have heard some terms related to machine learning, such as supervised learning, reinforcement learning, etc. In fact, these describe different training methods in the process of machine learning and are often applicable to different situations.
For example, a researcher wants to let an algorithm learn to recognize cats and dogs. If a large number of pictures of cats and dogs are input to the algorithm in advance, and it is told whether the picture is a cat or a dog, then this is supervised learning (Supervised Learning); if the researcher Give the algorithm a lot of pictures of cats and dogs, but don't tell it which are cats and which are dogs, and let the algorithm automatically find the difference between cats and dogs, this is unsupervised learning (Unsupervised Learning); if you let the algorithm keep doing multiple-choice questions , let the algorithm "look at the picture" and choose whether it is a cat or a dog every time. The correct answer will be rewarded with extra points, and the wrong answer will be punished with deduction. The algorithm tries to score as much as possible and avoid the deduction. The ability to recognize cats and dogs is Reinforcement Learning.
It is through training with a large amount of data that the machine has acquired such a powerful ability that even if the scientists behind the artificial intelligence "Alpha Go" (Alpha Go) are not master Go masters, and some even can't play Go at all, they can make it beat humans. Go world champion, which is impossible on a traditional machine. All the "behavior" of a traditional machine is programmed in advance by the designer, so it cannot implement behavior beyond the designer's understanding.
Machine learning is not invulnerable
Although machine learning makes many objects around people more and more intelligent, there is still a big gap from the real "strong artificial intelligence", because the current data-based AI algorithms have great limitations in many cases. For example, an AI algorithm that has been trained to "recognize cats and dogs" for a long time may mistake a Chihuahua for a cat, or a hairless cat for a dog, mainly because the results obtained by machine learning are highly correlated with the training data. If the data used for training AI is biased, for example, 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 algorithm obtained in this way can easily identify certain Mistakes when it comes to other kinds of cats and dogs.
Currently, the problem with AI systems is the interpretability flaw of machine learning, that is, machine learning is a "black box" process, and we cannot explain exactly what features it bases on to make judgments. When humans learn to identify cats and dogs, they tend to focus their judgment features on certain key parts of cats and dogs. An AI algorithm obtained through image training, even if the result has a high accuracy rate, may place some judgment features on the environment, which is obviously unreasonable and may lead to potential risks in its system application. For example, when developing an artificial intelligence self-driving system, if the developer cannot determine what it is based on to make driving decisions, even if the system performs perfectly in pre-launch testing, it may make fatal errors when encountering difficult road conditions. . There was a car accident in the United States, because the artificial intelligence self-driving system mistakenly recognized the white body of the truck as the sky, causing the car to hit it.
This also clarifies a truth from another perspective: although the sea-questioning strategy is useful, it is not efficient and will lead to potential mistakes. If you want to learn new knowledge, you still need to use causal logic to fundamentally understand the ins and outs of things. Currently, scientists hope to achieve this in artificial intelligence.
To achieve causal logic, machines still need to work hard
The founder of Bayesian Networks and Turing Award winner Judea Pearl believes that the key to making artificial intelligence realize the essential leap is hidden in each person's brain, that is, the powerful weapon given to mankind by God - causal logic.
Pearl divides thinking into three levels: the first level is association, which corresponds to the ability to observe, which is the level of current data-based "weak artificial intelligence"; the second level is intervention, which is associated with Correspondingly is the ability of the control variable to implement actions, that is, the ability to obtain cognition through intervention; the third level is counterfactual, which corresponds to the ability to imagine. Fortunately, the human brain is at the third level, and the imagination gives us the ability to construct counterfactuals - fictional worlds through imagination, and thereby construct cognition. For example, Einstein extended special relativity to non-inertial frames with acceleration through thought experiments.
The difference between association and causation is that association only reflects the most superficial information between data, that is, correlation. For example, there is data showing that there is a correlation between temperature and crime rates, with crime rates higher when temperatures are cooler. The crime rate during the Spring Festival in my country will indeed increase, mainly because the activities of thieves become more frequent during the Spring Festival, and it happens that the Spring Festival is generally the time period with the lowest temperature of the year. If we input the data into an artificial intelligence system that only knows how to analyze the correlation, and analyze the data only from the perspective of correlation, we will get the conclusion that "the crime rate increases due to lower temperature". If this artificial intelligence system is used to predict the crime rate of a country without Spring Festival culture, or the crime rate of a year with abnormal temperature, it is very likely to get a "nonsense" conclusion.
From the perspective of causality, researchers must not only analyze the correlation between the data, but also judge the internal logic chain. For example, when the temperature remains the same throughout the year, will the crime rate change? If the answer is "yes", then it can be assumed that there are other factors besides temperature. For example, due to the outbreak of the new crown pneumonia epidemic, the flow of people during the Spring Festival in 2020 decreased. Although the temperature in winter continued to drop as usual, the crime rate did not change accordingly.
Pearl believes that an important channel to rise from machine learning to causal learning is the introduction of "intervention" operators. There is a fundamental difference between "intervention" and "observation," for example, observing a rooster crowing and forcing a rooster to crowing are two completely different things. Current AI algorithms can determine the correlation between the rooster crowing and the rising of the sun, but it is difficult to determine whether the sun will also rise when the rooster is forced to crowing. Pearl believes that an AI that only accepts passive observational data cannot climb to the second level to answer questions related to "intervention" and understand the cause-and-effect relationship between "rooster crowing" and "sun rising". , because confirmation of causality requires control-variable experiments that are themselves intervention-based.
You may ask, if the observed dimension is rich enough and the data obtained is sufficient, can observation be a substitute for intervention? In fact, it is difficult for us to ensure that the data range is consistent with the actual test environment, and even more difficult is that it is often impossible to know whether the data itself is complete. This leads to the possibility that no matter how huge amount of data is used to train AI algorithms, errors may occur because the data is not completely consistent with the test environment, which is called the OOD (Out of Distribution) problem. Turing Award winner Joshua Bengio also believes that OOD generalization is one of the most urgent problems in current artificial intelligence.
Now, the common artificial intelligence systems in life are still far from "strong artificial intelligence". They do not have the ability to judge cause and effect, and they do not have the imagination of the third level at all. However, scientists have realized that "causal learning" is the key to the next leap in artificial intelligence, and many scientists have successively invested in the theoretical research of "causal machines". For example, Professor Cui Peng of Tsinghua University proposed to improve the OOD generalization problem by combining causal inference with machine learning (Stable Learning); More accurate predictions have been achieved on the non-stationary data of . With the development of technology, artificial intelligence technology will definitely become more and more reliable, and can ultimately benefit mankind.
.jpg)