Correspondence with Dahua Lin

第一封 我给林达华写的 Hi, Dahua, I am a junior college student majoring in EE from China. I have written you email before. I am interested in computer vision and machine learning. Currently, I am reading Bishop's book and reading papers on object recognition. Meanwhile, I have some thinking on computer vision and artificial intelligence and would like to hear your comments. Since machines are intended for some specific applications, the ultimate form of computer vision may not have too much resemblance of human vision. For example, to incorporate prior knowledge is more efficient than to let the machine learn from scratch. That is may be why discriminative models have generally better results than generative models while the latter seems more similar to the real form of intelligence. So, data-driven methodology will be widely used in computer vision, which is more concerned with real world, specific problems. Meanwhile, it is worthwhile to explore the universal intelligent mechanism, such as a global ontology or general learning, just for the reason that deeper understanding of intelligence itself can be gained. Mathematics plays an important role in backing up machine learning and computer vision while frontier problems are never solved by current math. One person has commented that "The limitation of artificial intelligence lies in the limitation of formalization". However, I think while it is true that present theories and technologies are constrained to formalized objects/concepts, the limit of formalism is evolving along the time. Maybe the following would better illustrate the picture. New problems are encountered by trials and heuristic approaches, then some formal methods such as algorithms are developed. At last, a new theory of math, stimulated by the new challenge and revising existing heuristic techniques, comes into being and solves the problem elegantly. Therefore, in exploring frontier problems, we should not be limited by existing math, but try to get a working solution even though it may be rigorous. I believe a key factor to learning is that the learner should be initiative. Personal experiences and studies from psychology have confirmed that learning is most effective in trying something actively. So, if machines are able to try besides sitting there to learn, a new level of intelligence will be achieved. But, a difficult problem lies in how to generate goals for machines. That is why no real active learning is available. I am considering object categorization problem at present. While models that capture appearance, spatial, texture and context information sound more natural, they do not perform as good as bag of feature models that are essentially descriptor histograms. Maybe the problem lies in a global model. Different objects should be perceived in different scales. For example, trees are usually seen in meters but leaves exhibit most dominant features in centimeters. And, obviously, in different scales, things possess different properties (e.g. nanometer-level world looks totally strange) and varied structures. Thus, a combination of diversified models should perform better than a fixed one. Besides initiative, another important characteristic of learning should be dynamic. In static settings, the ability of learning has a stronger limitation. This idea comes from my observation that even people sometimes face problems as "over-fitting." When one constantly receives a particular emphasis on, for instance, diligence, from people around him, he may be too tuned to that particular idea so that to ignore other important qualities such as creativity since "creativity data" are rare in the "training set." From this perspective, on-line learning might be a promising method in addition to its advantage for large-scale learning. I really appreciate it if you can give me some feedback. Thank you very much. Cheers.

大牛的回复 Hi, William, Frankly speaking, your understanding on vision and learning is really impressive considering that you remain a junior college student. I am confident that you will have great achievement in this field if you keep your persistence in academic exploration. In order to keep the words concise, I directly add the comments below each paragraph, such that I do not need to restate the context. ----------------------------------------------------------- William wrote: > > Hi, Dahua, > > I am a junior college student majoring in EE from China. I have written you email before. I am interested in computer vision and machine learning. Currently, I am reading Bishop's book and reading papers on object recognition. Meanwhile, I have some thinking on computer vision and artificial intelligence and would like to hear your comments > > Since machines are intended for some specific applications, the ultimate form of computer vision may not have too much resemblance of human vision. For example, to incorporate prior knowledge is more efficient than to let the machine learn from scratch. That is may be why discriminative models have generally better results than generative models while the latter seems more similar to the real form of intelligence. So, data-driven methodology will be widely used in computer vision, which is more concerned with real world, specific problems. Meanwhile, it is worthwhile to explore the universal intelligent mechanism, such as a global ontology or general learning, just for the reason that deeper understanding of intelligence itself can be gained It has been widely acknowledged that data-driven approaches work effectively in particular domains that they are specially tailored for, which are however at the expense of generality. It is our ultimate goal to achieve "universal" machine intelligence like that of human-beings, and the efforts devoted to this goal have never ceased. However, the current advancement of artificiall intelligence is still in a baby-stage that is very far from the destination. The debate between generative-discriminative models has lasted for decades. It is really an interesting issue. However, please keep in mind that they by no means represent the whole story of intelligence. They are just two forms of constructing statistical models. Though generative models look more natural in the perspective of statistics, it remains unclear whether it is the right formalism to capture the way that we learn and think. In recent years, statistical methods gradually dominate the research of machine learning, and thus the fields thereon. There's no denying that they achieve great success in various domains. However, their potency and elegancy in formulating problems do not serve as evidence that they are the unique correct direction of AI. In my opinion, a true-sense insight into the essence of AI requires exploration in more directions, in addition to statistics. > > Mathematics plays an important role in backing up machine learning and computer vision while frontier problems are never solved by current math. One person has commented that "The limitation of artificial intelligence lies in the limitation of formalization". However, I think while it is true that present theories and technologies are constrained to formalized objects/concepts, the limit of formalism is evolving along the time. Maybe the following would better illustrate the picture. New problems are encountered by trials and heuristic approaches, then some formal methods such as algorithms are developed. At last, a new theory of math, stimulated by the new challenge and revising existing heuristic techniques, comes into being and solves the problem elegantly. Therefore, in exploring frontier problems, we should not be limited by existing math, but try to get a working solution even though it may be rigorous. I agree on your understanding on how maths and the applied fields (like learning and vision etc) evolve during a long period. In general, they have mutual effects on each other: applications present new problems to stimulate the advancement of maths, which in turn offers new tools in solving the real problems. However, in a relatively short period, things do not necessarily go this way (at least in the fields of learing and vision), due to the lack of communication between mathematicians and applied scientists. In the past century, mathematicians have developed a variety of powerful tools to describe the world. However, most of them are formidable to a typical researcher in learning and vision. Only a very small group of researchers can understand modern maths, and even fewer can apply them to their research topics. As you read through the papers in learning and vision, many of them only utilize elementary tools developped a hundred years ago. Many of the new methods actually exist in maths and other fields (like physics) for quite a while, we are reinventing the wheels from time to time without awareness of their existence. In my view, research on AI problems should not be limited by the math tools that are available to us. However, it is absolutely beneficial to know more about the development of modern mathematics, in which you can acquire a lot of significant notions and methodologies that you may not even come across in AI papers. > > > I believe a key factor to learning is that the learner should be initiative. Personal experiences and studies from psychology have confirmed that learning is most effective in trying something actively. So, if machines are able to try besides sitting there to learn, a new level of intelligence will be achieved. But, a difficult problem lies in how to generate goals for machines. That is why no real active learning is available. > > > > I am considering object categorization problem at present. While models that capture appearance, spatial, texture and context information sound more natural, they do not perform as good as bag of feature models that are essentially descriptor histograms. Maybe the problem lies in a global model. Different objects should be perceived in different scales. For example, trees are usually seen in meters but leaves exhibit most dominant features in centimeters. And, obviously, in different scales, things possess different properties (e.g. nanometer-level world looks totally strange) and varied structures. Thus, a combination of diversified models should perform better than a fixed one. > > > > Besides initiative, another important characteristic of learning should be dynamic. In static settings, the ability of learning has a stronger limitation. This idea comes from my observation that even people sometimes face problems as "over-fitting." When one constantly receives a particular emphasis on, for instance, diligence, from people around him, he may be too tuned to that particular idea so that to ignore other important qualities such as creativity since "creativity data" are rare in the "training set." From this perspective, on-line learning might be a promising method in addition to its advantage for large-scale learning. Learning in an initiative way, incorporating perspectives in multiple scales, as well as continous learning, are all good ideas in improving an AI system. Actually, all these are active topics. You may refer to topics in active learning, reinforcement learning, multiscale, online learning, and evolutionary learning. Many of the methods developed in these topics adopt the strategies that you have mentioned in some manner. Though extensively studied, these topics remain premature in my opinion. There is still no consensus on how a learner should learn things initiatively and how information from different scales should be combined. In most works, systems are constructed in an ad-hoc manner, and tailored to a specific problem. I believe that mathematics may play an important role in unifying these works with a solid foundation, and giving guidence of further development. > > > > I really appreciate it if you can give me some feedback. Thank you very much. > > Cheers. > > > -- > Persistent efforts will be rewarded at last The comments are just my opinions, which may be correct or not. Learning from other's thoughts is pretty important in the way of research, however, creating your own way based on what you learn is paramount as well. Keep learning, thinking, and practicing, you will attain success. Best, Dahua

相当感动 又发了一封 Hi, Dahua, Thank you for your comprehensive comments. And for your encouragements. I totally agree with your view that "we are reinventing the wheels from time to time without awareness of their existence." From the blog entries, I have observed that you have found many powerful tools in mathematics which are out of sight to most of us. I do find current methods are in the baby-stage, especially in computer vision. Maybe it is due to its innate characteristics, computer vision, at the present, is mainly about inventing tricks that can work in certain areas. Indeed, far from a satisfactory solution. I am curious about your remark that there is something more outside statistics. In my limited knowledge realm, the only tool that can deal with uncertainty well is statistics and probability. Although I believe there must be better ways to explore deeper, (just as the wonderful calculus opened us eyes for a totally new world and rendered difficult problems in the past a simple one) is there any at present? Maybe this is the time when we try out all possible ways, and eventually, one inspiration will evolve to the "calculus" for tomorrow. I have some other question; it's about the research I am trying to do. I intend to mark all the correctly and misclassified images and descriptors by running existing algorithms. Then, I hope through some analysis into the results (by simple statistics, e.g. the histogram of error reasons), I may be able to find out some cues. However, I am concerning about how this approach will work. Since I have no clear idea of all the possible factors, I maybe focused on a false one exhibited in the data. This concern raises from my observation about statistics in social sciences. It is reported that almost every survey claims they have found some underlying rules and it fits well with their data and analysis. Unfortunately, different research, though reasonable from its own perspective, conflicts with each other. Therefore, I wonder in what a way can we track some rules, especially when many factors are present and their relationship is complex. I know you have sufficient experience in doing research and setting up experiments. In your view, is this method feasible? Thank you again for your reply. I am really encouraged. Best wishes.

结果又收到一篇很详细的回复 激动中 Hi, William, With respect to the issue of mathematical tools, linear algebra combined with multivariate statistics has been sufficient in developing most of the techniques that you have seen in papers. However, I don't think they are enough in solving the fundamental problems in vision and learning. You are right that statistics is the tool for modeling uncertainty. However, uncertainty is not the only thing that we need to deal with in our research. Before we talk about uncertainty and probabilistic models, we need machinary to describe, represent, and transform real world objects. Currently, most of the works treat them as they are in a finite-dimensional Euclidean space. Admittedly, it works pretty well in a small and constrained domain. However, when we come to a more general problem, do you think Euclidean space is the most appropriate space to put the objects in? In other words, statistical models are somewhat high level models, which are actually defined on some algebraic and geometric (topological) context. With the representations in such a context, we can then formulate and manipulate uncertainty. However, representation of real world objects, especially the vision objects, in itself is a fundamental issue yet to be solved. Some more advanced math notions may play a role here. Moreover, the application of maths is by no means restricted in solving the representation problem, they can also be applied to study how objects are transformed, related, etc. Let's come back to your research problem. Distilling entangled factors from their behavior is an interesting yet difficult problems. There's no panacea to problems of this kind. The solution varies from case to case. In my experience, to derive a good solution, a couple of things are required. First, reading through a series of papers to have a general ideas of what strategies are available and what are their pros and cons. Second, careful examination and analysis of the problem in your hand. Both are indispensable. My suggestion is that you first give a comprehensive analysis of the problem, and try to figure out whether there are some interesting observations. This may take a while but it will be rewarding. After you pinpoint the key issue, you can then select a proper technique to address it. You should be careful to keep yourself from being too much enthusiastic towards some particular methods or models that look beautiful. To a specific problem, careful analysis and observation is the key to the true solution. Best, Dahua

楼主相当积极啊

赞呀，多跟牛牛们交流。

真不错啊

嘿嘿 发现要常和大牛交流 提升会比较快 【 在 river 的大作中提到: 】 : 楼主相当积极啊

嗯 经常和蛋白质大牛交流 秀才 【 在 cryppie 的大作中提到: 】 : 赞呀，多跟牛牛们交流。