In this issue of Red Bird Salon, we are honored to invite Professor Xiong Hui, Associate Vice President of Hong Kong University of Science and Technology (Guangzhou) and former Vice President of Baidu Research Institute, as the keynote speaker, Professor Wang Yang, Vice President of Hong Kong University of Science and Technology, and Senior Vice President of Baidu and Director of Baidu Cloud Business Mr. Liu Hui, Chang , is our guest speaker. Today's Red Bird Salon discussion revolves around "Humanity and Sociality in Artificial Intelligence Algorithms" .
In recent years, artificial intelligence and machine learning have been changing society more and more, but artificial intelligence is often a black box in our eyes. How do we understand how artificial intelligence works? Are there similarities between the nature of artificial intelligence and human nature? Can human nature and sociality be used to understand and improve artificial intelligence algorithms? Even if you let the machine learn, what if the machine gets tired of learning? So what if the machine learning is involuntary? Let us share the humanity and sociality in artificial intelligence algorithms.
Red Bird Salon Issue 15 (Part 1)
Red Bird Salon Issue 15 (Part 2)
The Tao in the algorithm follows nature
When I observed all the algorithms from a more systematic macro perspective, I used the theory of the Book of Changes to sort them out, which can be divided into difficult, simple and variable. The same is true for adversarial generative learning algorithms. This algorithm has many applications, and automatic image generation is one of them. For example, the images you see are all celebrity faces recognized by the public, but none of them are real people. These faces are learned from the faces of existing stars, and then automatically generated faces that conform to the popular aesthetic concepts of the public. image.
Another function of it is Face Agent (an artificial intelligence model, usually used for face recognition tasks), which uses a photo of a child to calculate the child's appearance through an algorithm. In practical applications, this algorithm can be used to find missing children. As long as you have photos of missing children, you can use the algorithm to calculate what the child looks like at different ages.
Before explaining the social manifestation of human beings, we first introduce two learning concepts of artificial intelligence.
The first concept is supervised learning . There are two types of supervised learning: differential and generative.
A) Differential learning refers to classifying two things in two types of things with linear and nonlinear boundaries. For example, if a group of Koreans is chatting with a group of Japanese, we can quickly judge that one group speaks Japanese and the other group speaks Korean. The differential learning method is to judge the relationship between yes or no, 0 or 1, so differential learning is a method that can be practiced quickly.
B) Generative learning is a way of learning from start to finish. For example, we learn Korean from the beginning to the end, and then learn Japanese all over again. After learning, we know whether they speak Korean or Japanese, and we can know the content of their speech. However, this learning process is very time-consuming and requires a lot of energy and resources.
The second concept is confrontational generative learning . This algorithm combines the above two learning methods of differential and generative.
This method is to quickly determine the category that needs to be learned in a differential manner, and then perform generative learning to form a lifelong confrontational generative learning process.
For example, the United States sanctioned China and prevented China from using GPS, so China made its own Beidou system. The United States continues to sanction China's key areas, and China continues to invest resources, time and energy to make up for its weaknesses. This combines the advantages of the above two learning methods.
Whether it is from the level of human beings or countries, adversarial generative learning is in line with the laws of nature - Tao follows nature, and learning from nature is the most efficient learning method. In terms of algorithms, the algorithms grown in nature are also the most efficient, which is the so-called Tao follows nature.
The Three Meanings of the Book of Changes: "Difficult", "Easy", "Variable" and Algorithmic Prediction Model
"Not easy" supervised learning
"Difficulty" is the unchanging essence of the world, and this essence is that things will not change due to changes in time and place. The nature of these regularities is the foundation of the algorithm model, so building a predictive model requires grasping "difficulty".
Machine learning artificial intelligence algorithms make predictions based on the fundamental distinction of "difficulty" in everything in this world. For example, apples and bananas have the root of "difficulty" in both shape and characteristics. The supervised learning method grasps the "difficult" characteristics of the predicted object and builds a prediction model, so that the machine can distinguish between apples and bananas from "difficult".
In order to understand "difficult" , we need to rely on "simple" . Although everything in the world is changing rapidly and constantly evolving, the evolution must conform to the law. If you explore "easy", you can find "difficult" .
Convolutional Neural Networks(CNN)--Convolutional Neural Networks
Convolutional neural network subdivides things, finds out the root of "difficulty", and establishes a predictive model.
For example, the "X" written by different people is different, but if you convert "X" into a matrix and digitize it, you will find the root of "difficulty". As shown in the figure below, the red box in "X" is the root of "difficult", and each "X" will have such a structure.
Capturing "difficulty" through convolutional neural networks. The process of convolution is to strengthen, reduce missing data or remove noise, and then build a model to make the characteristics obvious. This is a "simple" process that highlights the "difficult" at different levels.
The entire algorithm process captures "difficulty" through convolution, and then paves the way on another level, convolutes again, and then captures "difficulty" on a new level. Through this process, enough feature images can be extracted, and then through the activation function (SoftMax), the result is "yes/no".
From the avenue to the simplicity, discard the dross and select the essence
Much of today's technology is "simple". Clustering (Clustering) is an unsupervised learning method, which is to classify similarities into the same group. For example, a bunch of fruits are divided into oranges, apples, and bananas according to their appearance characteristics.
There are two types of clustering, one is Partition, which divides things into different types, and the other is Hierarchical. Hierarchical classification is to put things together upwards, absolutely freely, and purely in a tacit way. However, because this method is too free and has no pre-set or prerequisite conditions, the mistakes in the early stage will continue to accumulate and magnify, and then produce extreme monopoly, which is difficult to correct.
Clustering is a kind of "simpleness". As a summary, it helps to grasp and understand "difficulty", and then provides assistance for supervised learning. However, it should be noted that the process of "simpleness" is easy to cause information loss and some errors.
Algorithm of "Change"
"Change" means that all changes are exchanged, but they all follow certain rules and laws.
First introduce two concepts, exploration (Exploration) and development (Exploitation).
Exploration is an outreach, going out and taking risks and trying different things, and finding better resources and situations.
Exploitation is an internal one, doing what you are good at in a comfort zone.
Nowadays, society often talks about involution. The reason for involution is that everyone likes to stay in their comfort zone and do what they are good at. As time goes by, the learning and work system gradually becomes more and more closed, but If everyone has the courage to extend outwards and make crossovers continuously, it can help overcome involution.
Reinforcement Learning
Reinforcement learning algorithm is an algorithm that can take into account both exploration and development, making a balance between internal and external. The essence of this algorithm is to "learn from correct mistakes", that is, to continuously make dynamic progress in making mistakes and correcting mistakes. This is a human-like way of thinking and learning. From error to correction, if the efficiency of convergence is high, the learning ability will be strong. In a closed system, involution will inevitably occur, so keeping the system open can achieve a balance between exploration (Exploration) and development (Exploitation), and can continue to innovate.
There are two entities in the reinforcement learning algorithm: agent (Agent) and environment (Environment). The algorithm first passes the feedback of feeling the environment, whether it is friendly or hostile, and then takes the next step. If the feedback from the environment is friendly, the agent will also take friendly actions, so that the reward (Reward) will be generated, and a new state will be formed, and the new state can further receive rewards. Of course, there are also negative rewards, so the agent will adjust and correct errors and form a new state. In this way, the agent and the environment will form a circular connection.
Algorithmic golden mean
The algorithm process requires a golden mean parameter (Golden Mean), which is used to evaluate the quality of the algorithm, but this "golden mean" parameter is very difficult to adjust. For example, a low error (Low Bias), low variance (Low Variance), can be between "overfitting" (Overfitting) and "underfitting" (Underfitting), then you can get an organic "middle way" , which is a good algorithm. From an algorithmic point of view, extracting these technologies and things in life is already a "middle way".
02
Guest interaction
Q
Mr. Liu : Peer pressure is the motivation among peers to make people learn faster and better. In the process of competing on the same stage, learn the method of merit learning, and find out excellent talents or managers through supervised learning (Supervise Learning). Another way is to find a superior person, a noble person, or an enemy (competitor). Learn advantages from superiors and improve yourself; strive for progress and change from the same competition with competitors. Seek different ways to help grow at different stages, constantly learn the current key points, and achieve targeted in-depth learning to maximize efficiency.
A
Professor Xiong : In the process of human growth, to find superiors, nobles, and enemies, different people are needed at different stages. A noble person is a good mentor; a superior person is a good role model; an enemy is a competitor and can help you grow. Let yourself learn the key points of the moment, learn effectively and in-depth in a targeted manner, and maximize your efficiency.
Q
Professor Wang : Talking about a social issue from machine learning— involved KPI ( key performance indicator). The characteristic of the Chinese is to use their ingenuity in the rules, and then make the KPIs the best, but the problem that this brings is that innovation is often kidnapped by KPIs, including in machine learning. To achieve diversification, there is a concept of "stochastic gradient descent". The original "middle way" has been changed to "stochastic gradient descent". Going down, the final training result will be more stable.
A
Professor Xiong : The concept of KPI is simply goal optimization. The KPI method is to assign tasks to you from top to bottom and determine the amount of work that should be completed. And OKR (objectives and key results: achieving goals through specific and measurable actions) is a bottom-up approach, where individuals set boundaries and then set indicators, which may more effectively stimulate creativity and Vitality, but this method may also change the setting of the algorithm's objective function. Whether the method is feasible remains to be tested. But if you can really combine the advantages of OKR and KPI at the same time , and set up a better optimization model, it is possible to design a better reinforcement learning algorithm.
Q
Mr. Liu : From the perspective of application, KPI and OKR have their own application advantages. For example, KPI guides the game. In a stable enterprise, things are easier to predict, so the KPI is not much different. On the contrary, in the case of strong subjective initiative or high variable innovation, the application value of OKR is relatively high, but it should be noted that OKR needs to maintain the core of adaptability, such as randomly generating surprises (Surprise) to encourage innovation, Avoid games and evolve into KPI variants.
From the perspective of social effects, fashion trends, and economic benefits, which of these three learning algorithms is more important now?
A
Professor Xiong : The three algorithms are actually very important. For example, reinforcement learning has many roles and applications in academia; supervised learning has produced many applications from traditional learning algorithms to deep learning (Deep Learning); simple applications are to help understand data and help judge performance. It is a natural and intuitive way to convert to supervised learning or semi-supervised learning, and then proceed to the next step of modeling. Therefore, from the perspective of fashion trends, reinforcement learning and deep learning are both hot spots. However, from the perspective of application, each algorithm has its own specific application scenarios and can also be combined with each other. For example, Alpha Go, which plays Go, uses reinforcement learning algorithms, but "simple" runs through the entire training and learning process.
02
audience question session
Q
Q : How to demonstrate ethics in artificial intelligence algorithms? How should ethics be quantified? Is this a parameter or an indicator? How does the algorithm work?
A
Professor Xiong : Part of the ethical issue is also related to use. For example, search engines will combine Obama and monkeys into search results due to the ranking of search volume, which leads to data objectivity and ethical issues. The data entered in the use of artificial intelligence is the key. Incorrect data can turn the model into a "monster". The chatbot "Microsoft Xiaobing" can become very violent because the chat conversations involve a lot of violence. Therefore, how to regulate the use of algorithms is the key issue, especially data input involves privacy issues. Now all countries will protect the security, reliability and privacy of data. I think ethics will be supervised and restricted more in the next step, because there are more and more appeals from various parties, but there is no standardized solution yet.
Q
Question : How should the people at the upper and lower levels communicate smoothly?
A
Professor Xiong : I think communication is a process of dynamic adjustment. Knowing how to reduce and increase dimensions at any time can make communication smooth and enjoyable. When chatting with others, you can judge the other party's knowledge structure and "knowledge tree" from the communication. If you are relatively shallow, you can reduce your own thoughts. On the contrary, if your "knowledge tree" is high, you can upgrade your thoughts. Because it will be difficult for the lower level to communicate with the upper level, and the communication between the upper level and the lower level will be boring, so proper dimension adjustment can make the communication smooth and enjoyable. Of course, you also need to have a certain level of knowledge to have room for dynamic adjustments. As the old saying goes, "the prime minister can pull a boat in his belly", if you are in the front of the account, then lower your posture a little. In this way, there will be no barriers to communicating with people on the upper or lower levels.
Q
Q : Artificial intelligence is good at summarizing a large amount of data, so can machine learning make independent inventions or find solutions to technical problems through a large amount of knowledge?
A
Professor Xiong : This is possible. The scientific discoveries of artificial intelligence may be easier to achieve than open topics (writing poems and essays). Because relatively speaking, the topic of a limited scene has a more objective standard and direction. For example, if we combine the layout experience of the predecessors with the knowledge of the structure of the semiconductor circuit board, it may be possible to automatically lay the cables through the method of artificial intelligence, but the premise is that enough data must be collected.
Click to read the original review exciting content
↓