Microsoft Asia Research Institute thanked: how to make the machine has the same thinking?

Can the machine enter the depths of human hearts to understand their character and emotions? These issues have been considered in the field of psychology for thousands of years. A few days ago, Microsoft Research Asia Senior Researcher Xie Xiong was invited to give a keynote speech "How to Make Machines Think Like People" on the theater-style speech platform. We have compiled thank you speeches and speeches. The full text is as follows.

Hello, everyone, I am thankful to Microsoft Research Asia. Today I want to discuss with you is, "How to make the machine has a human-like thinking."

Before I get to the point, I will first tell my own story. Recently, I bought my three-year-old daughter a picture book called "Can I build another me". She can't put it down. The protagonist of this book is a child who is tired of his regular life. He hopes to train a robot to take a nap, eat, and go to kindergarten on time instead of himself, so that he can play freely. So he bought the cheapest robot and brought it home to train it. In this process, the first question he encountered was how to make robots become him. So he tried to tell the robot about all kinds of information about himself, including his name, age, height, weight, parents, brothers and pets, and even the information of "left-handers", "easy to be annoyed," "the socks often have holes."

I found that the author of this picture book has a big brain hole. He is also thinking about the issues we think about. This story also tells us that the first step in getting robots to think about people is to understand yourself. Because we can tell robots how to do the most like ourselves. Today, I will discuss this issue with you from the following aspects:

1. Artificial Intelligence and Psychology

2. Personality classification and speculation

3. How to make the robot think like a human

For a long time, our team has been engaged in the study of user portraits. What is a user portrait? Simply put, it is through user-generated big data to guess and understand a person's age, occupation, interests and hobbies, as well as to describe a group of people's life patterns and movement patterns. This allows us to start thinking. Can we use these data to go deeper into people's hearts to understand their character and emotions? This is not easy. But in the course of research, we found that these problems have been considered in the field of psychology for thousands of years. In fact, the two domains of artificial intelligence and psychology have actually crossed.

Herbert A. Simon, one of the early pioneers of artificial intelligence, is a famous cross-border scholar. He is a computer scientist, a psychologist, an economist, a sociologist, and even a cognitive scientist. What is amazing is that he has achieved the same excellent results in every field: he won the Turing Award in 1975, the Nobel Economics Prize in 1978, the National Medal of Science in 1986, and the United States in 1993. Psychological Society Lifetime Achievement Award. On the right is Professor Geoffrey E. Hinton of the University of Toronto, an active facilitator of deep learning. He is both a computational scientist and a psychologist.

Two years ago, we began to visit famous psychologists and professors and tried to conduct interdisciplinary cooperation and exchanges. In this process, the first problem we want to solve is personality. Can human personality be calculated from user-generated big data?

Although the term personality is common in everyday life, it is not easy to give an accurate and clear definition of personality. Even psychologists find it difficult to reach a consensus on the definition of this term. The earliest definition of personality dates back more than 2,000 years ago (400 BC) to the body fluid of Hippocrates, an ancient Greek medical scientist, who believed that the human body consists of four body fluids, including blood, mucus, yellow bile, and black. Bile, and the distribution of these four body fluids determines the person's personality: black bile produces a melancholic personality, the blood has an optimistic personality, yellow bile produces impulsive irritability, and mucus produces a calm personality. Although Hippocrates' body fluids have been rejected by modern medicine, his exploration of personality classification is instructive, so that later psychologists have been discussing this issue.

When we communicated with psychologists, we found another interesting fact: In modern psychology, the definition of personality actually has a close relationship with the use of language. In fact, in the field of computer science, we also have a lot of research on language. We call it "natural language understanding." In psychology, there is a concept called "lexical hypothesis." What is the lexical hypothesis? According to this hypothesis, we do not need to observe and study a variety of people to study personality. We can simply observe the related vocabulary in human language directly. For example, when you introduce a friend to me, you may describe him in a large section of speech: "He likes talking very much. When a lot of people are, he is particularly happy and he speaks so much. He hears him every time. , is a saying, "and so on. In fact, one word can summarize this paragraph: talkative. Therefore, psychologists decided to organize these descriptive words. If this vocabulary is not much, they can become the basis for establishing a classification system. Based on these observations, Allport and Odbert, the pioneers of personality theory, conducted a difficult and systematic investigation of English vocabulary in 1936. By looking at the dictionary, they discovered about 18,000 words in the four categories of personal qualities, temporary emotions or behaviors, and intelligence and talent, and further compiled more than 4,000 vocabulary words that describe their character. Although it seems that there are only a few 4,000, it is still very complicated for the entire user language. Under the assumption that when describing a person's character, if it is necessary to score the four thousand descriptive dimensions, how much work should be done. Therefore, they want to further reduce it. In the process, they found that there are actually some correlations between these words. For example, an outgoing person is usually more talkative and a calm person is usually more sensible, but he may also be more introverted. If these correlations can be located, more than 4,000 words can be further categorized on this basis.

In the past two decades, the definition of personality that the personality researcher pays most attention to and supports is the five-factor model, which is often referred to as “big five personality theory”. As shown in the figure, the Big Five personality includes five highly generalized personality factors: Extraversion, Conscientiousness, Neuroticism, Agreeableness, and Openness. There are also some subdivision qualities under each personality factor (such as whether the extraversion includes frequent participation in activities, whether it is enthusiastic, etc.). In this way, when you introduce a friend later, you can describe him as "a person who is more outgoing but not easy-going and may be more emotional." The way is simple, but the description is comprehensive. In fact, the sorting of these vocabularies and the generation of personality classification systems are mostly data-driven, and they have many very close links with computer science. Can we automatically calculate the user's big five personality? In fact, this is also possible.

In traditional personality measurement, psychologists often use interviews and questionnaires. This requires a lot of manpower, financial resources and time. The subjects are often limited to a scale of tens to hundreds of people, and it is impossible to achieve large scale. User's measurement. Many people in this room may have done psychological surveys. Generally speaking, there are more than 100 questions. I do not know how many people will fill out these hundreds of questions seriously. It may be that everyone is playing "three" all the way - one to five minutes to score a middle point. This result actually has no meaning. This work is indeed very troublesome, and many times the respondents themselves do not actually know how to play a few points. For example, is introverting one or two more? In fact, they are all very vague. However, there is also a method of personality measurement in psychology called behavior measurement, which is performed by observing individual behavior. The theoretical basis of behavioral measurement is the consistency of human behavior in personality theory. Since personality can explain the stable individual differences between individuals, the differences in individual behaviors are closely related to individual personality. Therefore, by observing individual behaviors, it becomes possible to predict personality. Just before computer technology was widely used, it was difficult for psychologists to collect enough behavioral data from users, so the lack of data led to behavior measurement not being widely adopted in traditional psychology.

However, in recent years, with the popularity of the Internet, smart phones, and various sensing devices, user behavior data has been widely collected, coupled with the advancement of artificial intelligence methods in modelling users, and methods of measuring personality through behavior data. In the intersection of computer and psychology, rapid development has been achieved. Our research work goes further and proposes a "personality inference model" that uses heterogeneous data on social media (such as portrait photos, published text, emoticons, and social connections) to predict Big Five personality. For example, for pictures, we can use the deep residual network to calculate the semantic representations and then group these pictures into certain categories, such as cartoons, selfies, group pictures, animals and plants. In fact, in the process, we still need to cooperate with psychologists. Using personality-based artificial intelligence methods based on behavioral data, we first need to collect a small number of users' questionnaires as annotations. By annotating user behavior characteristics and personality characteristics, the mapping and connection between them are input into the model to train a good model.

In fact, we found a group of volunteers who provided their own data and completed a questionnaire survey so that we have two data. After training the model, new users do not need to complete the user survey. The model can automatically calculate its personality. We can look at the calculation results. Does it sound abstract? However, it is actually very specific. For example, we can calculate the relationship between the user's published text and personality. The Big Five personality has five dimensions. We can calculate that the text is particularly positively or negatively related to each dimension. For example, a person who often writes about youth and self in a circle of friends may be more outgoing, while the user who often fails to write and faces extravagance will have very low scores. There are also some users who may write vocabulary such as age, society, and success that sound very positive. We find that these people are more responsible. On the contrary, some people may often write the words casually, swiftly, and temperamentally. We find that they are less responsible. Low conscientiousness is not a derogatory term: In this model, the person who cares about the result has a higher degree of conscientiousness, and the person who cares about the process is less conscientious. Both of these extremes have their advantages and are not good or bad.

We also calculated the Pearson coefficients of the big five personality and user avatar clusters to show clusters that are strongly positively or negatively related to Big Five (two clusters per cluster). Such calculations reveal some interesting phenomena: For example, users with high extroversion scores prefer to use avatars with smiling faces, while users with low scores often block facial expressions or use side faces in their avatars; users with high openness scores often use them. Photographs with friends as avatars, and many with a low openness score are self-portraits. Our experimental results show that the use of avatar photos alone can make the accuracy of individual personality predictions reach 0.6. We not only propose targeted feature extraction strategies for behavioral data in each dimension, but also use integrated learning technology (Ensemble) to effectively integrate different dimensions of behavioral data to improve the accuracy of Big Five personality prediction and make individual Big Five personality. The accuracy of the forecast reached 0.75 or more.

After understanding the user, the next step is how to use this knowledge to help the robot to produce human-like thinking. One of the important behaviors that humanity hopes robots can achieve is chat. Microsoft also proposed the concept of "Conversation as a Platform", in which all human-machine interfaces in the future will be transformed into dialogue interfaces.

I have seen a TV drama two years ago and still vividly remembered. It was the first episode of the second series of the Black Mirror "be right back." The TV series describes an artificial intelligence company that can synthesize a virtual person through a person's social media and online chat data to imitate the personality characteristics of the character prototype and engage in dialogue with his girlfriend. This may seem like science fiction, but it is actually not far away from us. In a news report in October 2016, Kuyda, an entrepreneur from Russia, to commemorate his dead friend Roman, trained a chat robot with his 8,000 SMS data and officially released it in May 2016.

Although technology has taken a big step forward, even the best chat bot currently cannot feel that he is a stable, emotional, and living person. This involves how to make the robot's language and behavior more personal.

With the prevalence of social networking, language data with user tags becomes easier to obtain. As described in the news report mentioned earlier, if we have enough data about a person, it is possible to train a chatbot that is the same as his personality. Of course, we can also train robots with a class of people through data from a group of people, such as children, students, and even poets. For example, can we collect data for all modern poets and use this data to train a robot that exports poets? Its implementation can also be done. However, with the deepening of research, I believe that in the end we will encounter bottlenecks, such as how to make robots have more realistic human characters and emotions, which still needs to cooperate with psychologists.

In fact, Eliza, the oldest chat robot, is a psychological counselor. About 50 years ago, a researcher at MIT, Joseph, developed Eliza. When chatting with users, Eliza introduced the personal-centered therapy proposed by psychologist Rogers, with more emphasis on dialogue, such as respect and understanding. Reason. Eliza actually does not take the initiative to say something new. It is more of a guide for users to talk as much as possible. The seemingly intriguing Eliza project was an unexpected success. Its effect shocked the user at the time, including its creator, Joseph. In fact, Joseph gave the project a name for ELIZA at the time. I do not know if you have seen "Flower Girl"? In this drama, Eliza is at the bottom of society. In order to enter the social upper class, she worked hard to learn the language used by the upper class people and made her look like an upper class person, but eventually the disguise was dismantled. Joseph named the robot ELIZA, hoping that the machine can disguise adults, but what he did not expect was that this disguise could not be easily dismantled. As a result, a vocabulary was later produced, called the ELIZA effect, which was a psychological feeling that overestimated the robot's ability. The implementation of the ELIZA effect is also very common. For example, when AlphaGo is defeated by top players, people think that the computer has the inspiration to go and the artificial intelligence will soon surpass humanity. But in fact, all the programs behind AlphaGo are written by people. The so-called inspiration, so-called intelligence, is ultimately achieved by the program.

Inspired by the ELIZA project, Microsoft Asia Research also launched the DiPsy project. The goal of this project is to allow robots to chat with people and help them overcome psychological problems. In this project, we drew on Cognitive Behavior Therapy and Mindfulness commonly used in psychological counseling. DiPsy's feature is to guide dialogues in a natural and effective way, allowing users to dedicate themselves. It will also study the user's psychological process and drive data-driven diagnosis of the user's psychological traits and mental disorders. We use cognitive behavioral therapy (CBT) or early intervention to change the way users think and behave in a variety of therapeutic contexts, helping users at risk mitigate and manage their psychological problems.

In the future, we expect this project will help solve practical social problems, such as the psychological counseling of rural left-behind children. At a recent forum held in the near future, Shen Xiangyang, executive vice president of Microsoft Global, said that he wanted to solve three diseases that are closely related to the human brain: childhood autism, middle-aged depression, Alzheimer's disease. I hope our technology can help him to do this.

Of course, many of these research projects are still in their infancy. We hope that we can eventually realize that the machine has human-like thinking and can provide not only help but also company when people need it. When you are alone, at least one AI is with you.

Flexible Pcb

What is flexible PCB used for?

A flexible printed circuit board is a PCB designed to meet the needs of flexible electronic circuits. The flexible printed circuit was originally designed to replace the traditional wiring harness. From its first application during World War II to the present, flexible circuit boards have increased exponentially.
Use flexible laminates to produce flexible printed circuits. The flexible laminate holds the conductive foil and the dielectric substrate. The flexible circuit board can be three-dimensional wiring, and its shape can be adjusted to fit the available space.

flexible pcb,flexible pcb manufacturing,flexible pcb connector,flexible pcb fabrication,flexible pcb board

HAODA ELECTRONIC CO.,LIMITED , https://www.pcbhdi.com