风之谷: Acdemic

About Valuable Academic Work

Fri, 30 Mar 2007 08:02:21 GMT

忧伤的小柿子 has a good comment about my former post Science 2.0 as follows:

"It is good to share scientific achievements more conviniently but here comes a question: if ppl publish academic results on their blogs or some forums how can they protect their copyright since it is so easy to copy and spread materials on those places? Besides, how could they be valued and graded?"

Since the answer might be related to some other friends, I'd like to reply it as a new post:

First, in my opinion, most of the published academic results are valueless.

I think only two kinds of work are really valuable. One is the "starting" work that set up a direction or area for others to pursue. The other is the "ending" work that totally solve the key problems of an interesting or important area.

For the first one, most of the people will acknowledge the contributions of the people from whom they have benefited. Note that, it might not be the one first conceived the idea but the one who has influenced the people with idea first. I wrote a paper about visual keyword at 2002. One or two year later some guys from CMU or Berkeley wrote some papers of the same idea. These guys are better at selling the ideas. Now most of the related CV people know about Video Google but not aware of my former work, hehe. I'm not complaining about it. It's just like people know Windows of MS or the OS of Mac but not know the Xerox PARC lab that invented the graphic UI of PC and Mac.

For the second one, if the solution could be monetized, it should not be published at all. Page did write a paper of PageRank given the fact that Kleinberg would like to publish a very similar work of HITS. But he never published any following work during the following decade. If the work could not be monetized in a short period of time, e.g. some mathmatical proofs, it should not be published either since the people who's job or life is betting on the problem might lose some important things:-). For example, since I've basically solved the image search result clustering problem, a guy at CMU has to change his Phd topic, hehe. Actually, lots of important theory work of mankind has been buried for several years. One of the reasons is that the real innovator of them proved their ideas for dreams not reputation.

Science 2.0

Mon, 05 Mar 2007 16:13:34 GMT

The New Science of Sharing from Business Week
Companies such as Novartis and Intel are at the forefront of Science 2.0 by encouraging open systems of collaboration

“Just as the Enlightenment ushered in a new organizational model of knowledge creation, the same technological and demographic forces that are turning the Web into a massive collaborative work space are helping to transform the realm of science into an increasingly open and collaborative endeavor. Yes, the Web was, in fact, invented as a way for scientists to share information. But advances in storage, bandwidth, software, and computing power are pushing collaboration to the next level. Call it Science 2.0.”

“Conventional paper-based scientific journals, meanwhile, will be augmented by dynamic publishing tools such as blogs, wikis, Web-enabled RSS feeds, and podcasts that turn scientific publications into living documents. Projects such as MIT's OpenWetWare are already doing this.”

“Intel established exploratory research labs adjacent to leading research centers such as Berkeley, Cambridge University in Britain, Carnegie Mellon, and the University of Washington. Intel provides the funding and each lab houses 20 Intel employees and 20 university researchers. While each lab has a unique focus—such as ubiquitous computing or distributed storage—the research teams from each lab meet regularly, as Intel has found that some of most promising insights and applications flow from unexpected synergies that arise when teams from different institutions discuss their research.

And rather than wrangle over who gets to control and exploit the fruits of joint research efforts, Intel and its academic partners sign Intel's open collaborative research agreement, which grants nonexclusive rights to all parties. Both sides retain their freedom to engage in further research, develop new products, and partner with other players. Like the pharmaceutical industry, Intel is finding that the benefits of casting a wide net for new ideas and learning rapidly from the external research ecosystem greatly outweigh the advantages gained from keeping the basic scientific research proprietary.”

In the mainland of China, Intel is collaborating with three universities: Peking University, Jiaotong University of Shanghai, and Fudan University.

Sooner or later, people will publish academic results on their blogs or some forums. I hope that day will come earlier for I hate the current reviewing and publishing progress. It is biased, unfair and a little bit waste of time:-).

ML, CV, IR or HCI?------ML

Tue, 26 Dec 2006 01:37:20 GMT

最近和一些朋友讨论研究方向的问题。我说说我对我比较熟悉的几个领域的一些个人看法吧。纯属个人看法！如果和你的想法不一致就当我在胡扯好了。这个东西没法辩论的。有点像和人讨论彼此老婆孰优孰劣一样，会打架的，呵呵。

我第一个接触的领域是机器学习(ML)。那个时候我大三。每天花不少时间泡在清华的图书馆里看相关的书和Journal。那个时候还没有Google，更没有Citeseer和Scholar。找东西真是痛苦呀。而且学生还不能把英文书借出来，只能每次去外文阅览室里面看。那时神经网络比较火。我也在跟张老师做些相关的东西。所以这方面的东西看的比较多。当时对做研究几乎没有概念。也没人教我。所以效率比较低。当时看的东西里面给我印象比较深的有几个。一个就是经典的BP算法，一个是SOM，还有一个比较神奇是VC维。记住这个可能多半和他的名字有关系，那个时候我偶尔要吃VC的，呵呵。大四的时候用了接近一个月的时间实现了张老师当时提出的一个前向神经网络算法（当时是第一次写几千行得code，写完这个就对写code这件事没有什么恐惧心理了。虽然现在我还是不喜欢写code这件事情。我还是喜欢用人类的沟通方式来进行沟通)。这算法和BP不同，是生构造出来的。连神经元都是必要的时候添加的。我对算法提出了一些改进的方法，也实现了其中的一部分。我的本科毕设就靠的这个。当时的模式识别试验也用的这个算法。本来我以为会很牛的。后来发现也就那么回事。毕竟只是个神经网络算法罢了。倒是当时想的一些feature起了不少作用。刚开始读研的时候，张老师发现了SVM并极力推荐给我们。于是我开始研究SVM。我深深的被SVM的简单之美所吸引。也知道了为什么神经网络不work。但是让我仍然觉得不爽的是他还是不能告诉我该怎么选模型定参数。虽然他号称是全局最优的。即使这样，SVM已经远远好与传统的方法了。尤其是在那些特征不是很好，数据分布也不是很规律的问题上，SVM往往都能取得很好的效果。比如后面要提到的基于内容的图像检索。我还记得我在组里的一个讨论会上和一个老师一起给大家介绍SVM。我提出我的疑问后，大家也没有什么反应，呵呵。后来来了微软后接触了Boosting。主要是Adaboost。我觉得这个方法很好，因为他比较scalable。我做研究的宗旨和一般人不大一样。很多人是要做最好的研究，不管有没有用，呵呵。我的宗旨是要做真正有用的研究，不管看起来是不是最好的，呵呵。我所谓的真正有用的一个标准就是这个东西有没有人愿意买，愿意花多少钱买。花的钱越多就越有用。Adaboost属于那种比较有用的研究，因为他比较实用。把本来做不到实时的人脸检测问题基本上做到实时了。想法也很简单，很容易理解。我记得我还特意回实验室和组里的同学老师分享这个好东西。再后来，就接触了一些Michael Jordan等人推行的Graphic Model相关的东西。打一开始我就不喜欢这类概率的方法。我的理由很简单。他们做了太多不切实际的假设，比如一些明明不是高斯分布的要假设成高斯分布，因为只有高斯分布才能解，呵呵。还有很多明明不独立的东西要假设成独立的，原因也是不做独立假设就没法解了。这样假设来假设去的结果就是解了个数学题而已，出来的东西基本没用。这个领域的人多半都是搞数学和统计出身的。数学基础都很好，至少做数学题的能力不错。我后悔当初做过得很多数学题没有留下来了，不然，每年攒几篇灌到NIPS应该没什么问题的，哈哈。机器学习就先说到这里吧。等讲到IR和HCI的时候我再回过头来谈谈如何能让机器真的看起来有些智能。

Two ICASSP papers were accepted!!!

Mon, 18 Dec 2006 16:52:45 GMT

I just got the notification letters of two ICASSP papers with me as a co-author. One is En Cheng's paper on relevance feedback. It is an extension of our work published in ACM MM 06. Congratulations to En for her 100% accept rate until now (She has three papers accepted by ICME, ACM MM and ICASSP). The other one is Richard's paper about MTV auto-generation. Both the idea and the direction are very interesting.

ICASSP will be hosted at Honolulu, Hawaii next April. Guo Bin and I will go there for a romantic trip:-).

A long version CHI paper about IGroup was finally accepted!!!

Mon, 18 Dec 2006 16:47:26 GMT

IGroup has been submitted to CHI last year. The meta-reviewer suggested me to segment it into two papers with one focused on algorithm and the other focused on UI aspect. Since it sounded very reasonable, I even did not rebuttal for it. I submited the algorithm paper to ACM MM 2006 as a full paper and a demo. Both were accepted. More importantly, I got to know Eldon through IGroup and VirtualTour:-). The UI paper was submitted again to CHI this year. The initial review results of the paper is very negative (2 in a 5-scale rating). Considering that there are near 2,000 papers submitted to CHI this year, it's almost impossible to change the decision back. However, Shuo and I felt a little unfair for most of the critiques of the paper were caused by misunderstanding. Therefore, we did a very solid rebuttal. Although, only 2000 words were permitted, we provided a link in it for further reference:-). The rebuttal did work this time. Our paper was finally accepted as a long version paper of CHI this year. I think the accept rate should be very low (should be less than 5%). This one should be the paper of mine with lowest accept rate. Shuo, Qixing, Jibo and Like, thank you for the great job:-). This proved my belief again: Impossible is Nothing!

对蔑视MSRA人的回复

Tue, 12 Dec 2006 07:22:09 GMT

Keso的Blog上有人对MSRA和IEEE Fellow很不屑的样子，还号称是这行的。在人家的地盘上我不好说什么。其实我觉得他那里留言的一些人真的是很不理智，很主观，也有些狂妄的。经常没有证据和理由的说些很不负责任的话。我真的不大想回复他们，不过觉得既然人家和你说话了，你总得回应一下表示一下基本的礼貌。所以我写了如下的留言。他的留言可以去这里看。

“Spirite,
你说的Google China和MSRA的不同我非常同意。这可能也是为什么他在MSRA上可以很成功，在Google China上一开始就不是很顺利的原因吧。陈华是我的朋友，他原来就座在我边上。他从MSRA学到了什么东西，只有他自己最清楚。我没权利评论。但是我想MSRA这段经历对他是有意义的。你可以自己和他交流。Kooxoo是很有用的东西，这个没有什么好争的。但是MSRA究竟在做什么你如果不清楚，请不要妄加评论。真正有用的东西是不会写成paper的。我一个Google的朋友说如果一个公司按照Google的paper做，那就一定完蛋了，呵呵。这个是很显然的事情，算法方面的东西别说paper了，就是patent都不要写。patent也是可以查到的。UI方面的东西倒是越早file patent越好。其实你看一个机构对公司的重要性是很容易的，你看他成长的速度就可以了。MSRA每年都要招相当多的人。去年招了近百人。如果不重要，你觉得Bill会傻到要花这么多钱雇这么多人吗。另外researcher的cost是要明显高于developer的，无论是工资还是travel的费用。Intel研究院的大量裁人就是一个典型的例子，呵呵。另外，你是IEEE Fellow吗？如果不是请等你拿到Fellow以后再评论吧。一个没有吃过包子的人评论包子好不好吃你觉得有意义吗，呵呵。”

Upcoming Conferences List

Tue, 07 Nov 2006 03:57:06 GMT

花了两天时间整理了一些接下来的和Multimedia, IR, Web, Data Mining, HCI相关的会议列表。放在我的主页上了。这期间仔细玩了一下Libra。搜了很多人。发现了一些很有趣的现象，知道了哪些人在灌水（很多论文缺没有citation)，哪些是真正的牛人（平均citation在10以上)。比如Andrew Yao和Sergin Brin就很牛。从影响力（学术上和工业上的）来看，后者更牛。我觉得下一届Turing Award可以考虑给Larry Page和Sergin Brin了。我自己的平均citation是1.5（16 papers，24 citations）。虽然这个比Scholar上的要低很多，但是整体趋势是一致的。而且每个citation的质量也更高。

感谢一把：

感谢Libra提供的conference排名！

感谢老婆的支持！

希望这个List对相关领域的朋友有帮助。

100% accept rate

Tue, 11 Jul 2006 05:56:24 GMT

刚刚收到ACM MM short paper的通知，我的三篇paper都中了。这样这次全部6篇paper就100%命中了。恭喜程恩，长虎在顶级的多媒体会议上有了自己的paper。大家提议如何庆祝吧，呵呵。

ACM MM06

Tue, 04 Jul 2006 08:41:14 GMT

这次ACM MM我作为author的一共有6篇，2篇full paper，3篇short paper，还有一篇demo paper。目前full paper和demo paper都有结果了，3篇全中了。两篇full paper都是application track，这个track的中稿率是16％（16 out of 99）。其中一篇关于image search result clustering （IGroup）的paper在被据了3次（SIGCHI, WWW, SIGIR）后终于被最适合的会录用了。这次如果再不中，我都不好意思再投了，呵呵。在此感谢当初为IGroup和EnjoyPhoto做user study的同学们。也要感谢DWight帮我们polish英文。希望short paper一样会有好结果。Cheers!