Subscribe cloud computing RSS CSDN home> cloud computing

Google Scientist AI talk about Principal: why the importance of knowledge

Published in07:43 2015-12-29| Time reading| sourceAcm CaCm| ZeroArticle comments| authorShoham Yoav

Abstract:Today's concern with the artificial intelligence and the 90's when there is a big difference. 20 years ago, the focus of the focus on artificial intelligence is based on the logic of AI, usually belong to the knowledge representation, namely KR, and today's focus is on machine learning and statistical algorithms. But the author thinks it's too late to lose something valuable.

Today, the focus of artificial intelligence (AI) and the 90's when there is a big difference. 20 years ago, the focus of the focus of artificial intelligence is based on the logic of AI, usually belong to the knowledge representation, that is, KR (Representation Knowledge), and today's focus is on machine learning and statistical algorithms. This change is very helpful for AI, because the machine learning and statistics to solve specific problems (such as image recognition) provides an effective algorithm, and KR has never reached this effect. But I think the pendulum has turned its head and lost something of value.

Knowledge representation is not a single content. If a complete overview of knowledge is to be expressed, I will focus on its "Applied Philosophy" - a logical representation of common sense concepts, which will focus on explicit semantics.

I use a personal story to illustrate. The story begins in 2009 I in the Journal of Philosophical Logic: a paper, then at Stanford University and Duke University to do research projects. Then do a company, Timeful. Finally to Timeful is acquired by Google for at the end of the story. The point of the story is that the original journal paper has a direct link between the company's eventual success.

The journal paper is called of Intention and the Database Perspective Logics. Prior to this, there was a brief but important paper on the intention (intention) logic AI, proposed by Cohen and Levesque creative, the paper named is Choice Intention + Commitment. The literature is inspired by the less formal philosophy of rational agents, such as Intentions, Plans, and, Practical, Reason, and Bratman. My own paper was inspired by Cohen and Levesque's paper, but questioned their basic philosophy, and proposed an alternative approach. Although my method is based on the calculation of power (as shown in the title), the arguments are purely theoretical and purely philosophical.

Following this journal article, I look for some money to continue to study, as the professor always does. And the donors tend to do is, they asked me to incorporate some of the potential applications of this work. And then something happened. The first is that I have been dealing with the intention - in my personal calendar. Second is that these are very special intentions - rigid events and conferences. Third is my personal schedule is not completely different from my grandfather, showing how people's time needs to change, and how science and technology progress, this is very strange. Eventually raises an obvious question: what happens if I use a more versatile and flexible intent type to improve the schedule, and the calendar has to deal with the complexity of the results?

For a better understanding of this point, it is worth to further discuss the concept of Journal papers. Database from the perspective of the proposed in the drawings, appended drawings may be to imagine is used to belief revision AGM belief revision) planning of generalization, the latter being limited in the picture "faith (belief)" part. In the AGM framework, intelligent database responsible not only for storage Planner (Planner) belief, is also responsible for maintaining their consistency. There are two types of databases in the enhanced framework, one store belief (beliefs), one memory intention (intentions), which is not only responsible for maintaining the consistency of each class of databases, but also maintaining the consistency between them. In the journal paper, I put forward the main consistency conditions, and in the subsequent papers, and Icard and Pacuit given its logical form, which is a conservative extension of the AGM framework. From this point of view, more technical details are not appropriate, and many of them do not. But it is important that the intent of the database in agent with intelligent functions.

Now go back to the main story, the sponsor was persuaded, we started a small project to explore these ideas. The next two years is very funny, but this story is not quite relevant, except: this project quickly by a new doctoral degree graduates Jacob bank leadership; at the same time my old friend and colleague Daniel Ariely is also involved in. He is a famous economist, in early 2013, we decided to start a company, finally said Timeful. Our joint research does not drive us too much, because it is not suitable to realize how urgent and how to deal with the problem in the society.

When the Timeful 1 came out in July 2014, the reaction of the users and the news media was very favorable. Received 2000 mail users in the first month, most of them are emotional. It has obviously touched the nerves of Timeful, even though the product still has a long way to go. Soon, the company attracted the interest of major players, and eventually was acquired by Google. If it is not KR, it is not possible, the following is the reason:

The object as the basic data model

Timeful developed a personal time assistant (PTA), and its role is to help manage the time, resources are the most scarce and the most difficult to manage. This approach is based on three main pillars. First, it allows the user to naturally express all the things that occupy the time in the system. Second, machine learning and other algorithms are inherently difficult to optimize. The third pillar is behavioral science, which means making an environment that helps to correct the time management mistakes we make (such as time delays and the high availability of the future). Among them, I would like to focus on explaining the first pillar; it is the most basic one of the three pillars, and is also directly based on the KR pillar.

Consider all things that will take up our time: meetings, events, errands, projects, hobbies, family, health maintenance, sports, or time for thinking and learning. Their surface is different, and in different application scenarios (meetings and activities are listed in the calendar, in the to-do list of errands. Items on the list in the project management system) or just to stay in the mind. They all take the same resources - time - if you want to make a sensible trade-off, they should be listed in the same place. In fact, they are all intentions (intentions), although they have different attributes. According to the idea of intelligent intent data, the first basic judgment is to develop a data model, enough to cover all types of intentions. The result of this data model is called the intent object (IO). IO is a character vector, including text, temporal attributes (when it can be performed, what time should be performed, duration by different degree of accuracy to indicate), executive conditions of intentions (such as location information or tools), and other property types.

Intent object has become the foundation of the system, all things, including scheduling algorithm and behavior nudge, rely on to them. Of course, the user does not appear as a feature vector, but with a number of pre packaged intent class put together. As of April 2015, a four: events, such as meetings); tasks, such as call); (habits habits, such as a week jogging three times); and projects (projects, such as writing a long report). But at the bottom, the system breaks them down into a feature vector.

More product decisions

Knowledge representation is not only the source of Timeful, but also the inspiration of its data model. When making specific product decisions, the team has repeatedly sought guidance from the philosophical literature. It is difficult to fully express this, here are two specific examples.

The first example, we must carry out a moderate mark.A list of all the to-do list allows you to check the task. Timeful also has this feature, but the trouble with us is that the task is marked but the event does not, even though they are all IO. This is not so much the aesthetic asymmetry, it is said that the underlying principle, and how these principles should be applied to other IO, such as habits and projects. Then we went back to the source and realized that we had to track the user's commitment (commitments). If there is a principle by which the philosophical literature agrees, then it is the intention to contain a commitment (reflected in the title of the Cohen and Levesque's thesis). When I'm going to do something, I don't just make a note of it. I'm determined to keep track of it and do it. From this angle, we realize the time do not need to track; a meeting was arranged it is complete (there are exceptions, such as if the meeting may have failed to reach the target, but these by the meeting of independent tasks). All other types of intent are required to be clearly monitored, so we end up with all the IO tags attached, in addition to the event.

When you build a product, you want it to have an inner beauty.

Second examples must be dealt with the intention of the time range.Most systems are "shame to-do list" - you down but never do. We want to avoid this situation, and to achieve the strict definition of the time range. This early decision goes back to a small debate in the literature. In the form of Cohen and Levesque, like "I intend to read this book," this statement is the basic concept of. But in my journal paper, I think this is a problem, it can be traced back to the problem of commitment. If I want to commit to a certain no fixed time, what is the intention of me in the end to what is the commitment, as well as it is how to really drive action? If you have a teenager at home, you know what I mean. On the contrary, I think that the basic structure should be similar to the "I intend to study at 2 pm on Saturday to 4 points," such a statement. Then you can relax with the presence of the existence, and give a purpose such as "I intend to spend two to three hours on the weekend". But you always have a clear time range. Timeful used to the idea; and the user's implicit contract is, she should seriously treat her intentions, in return, the system will help her to achieve the intention, by their inclusion in her schedule and to promote her finish (application startup's slogan is "get it scheduled. Get it done"). Therefore, each task requires a "on do" or "expected (by do)" date. Then, the task will appear in the time frame, next to the event. In the case of "by do", the system picks up the time before the deadline, and the user can modify it as needed. In fact, if the task is replaced at a later time, the system will automatically move the task. The same logic applies to habits and projects in a more complicated way.


The Timeful story is to the satisfaction of all, many parts are associated with KR. Is a person able to have the same opinion without KR or philosophy? May be, but in fact no one can do it, I think this is not an accident. When you build a product, you want it to have an inner beauty. What I mean is, usually, when you begin to design a great user experience, you have nothing to do or not to do it or not. If the internal structure is not correct, you will never get a real sense of the user experience. Philosophy and KR encourage you to think about the conceptual framework, and provide guidance when designing a specific function.

This does not reduce the importance of machine learning and statistics. But machine learning requires feature space, and statistics require event space. Even the most popular deep learning fans will not think that those who always produce mechanical Ji, independent of human insight (unless you work for Google, and only interested in catsa).

Does this mean that every philosophical puzzle and logical puzzle has a direct practical significance? Certainly not. But if you are designing a car, you really need a wheel, so you don't need to reinvent the wheel, especially if you don't have the wheel.

We have reason to be optimistic. There are indications that the researchers more and more views of machine learning and statistics will solve all problems "to skepticism, for example, recently AAAI Symposium brings together the knowledge representation, machine learning, leading researchers in the field of linguistics and neuroscience, to discuss the interaction between these domains. My feeling is that the pendulum has begun to swing back slightly, and if we act as a community to encourage this trend, AI will be better.


  1. Spring Symposium on KRR: Integrating Symbolic and Neural Approaches Stanford, University (Mar. 2015), AAAI (); Https://
  2. O n Alchourr, C.E., from G rdenfors, P., and Makinson, D. on the logic of theory: partial meet contraction and revision functions. Journal of symbolic logic 50 (1995), 510 - 530.
  3. Bratman, Intentions M., Plans, Practical Reason. Harvard University Press and, 1987
  4. Cohen, and Levesque P., Intention is choice H. commitment. J. Artificial Intelligence 42 (1990), 213 - 261
  5. Icard, T., Pacuit, e., and Shoham, Y. joint revision of belief and intention. In the joint of the proceedings of the 12th International Conference on principles of knowledge representation and reasoning (KR), 2010.
  6. Shoham, Logics of intention and the database J. perspective. Y. Philosophical Logic 38 (2009), 633 - 647

About the author:Shoham Yoav ( is Google's chief scientist, Professor Standford (retired).

Original address:Knowledge Representation Matters Why(Translator / reviser Liu Xiangyu / Zhao Yihua, Liu Diwei / commissioning editor Zhong Hao)

Translator introduction:Liu XiangyuIn the general software development engineer, attention, machine learning, neural network, pattern recognition.

step on