时间: 2014年7月2日-31日
参加人数: 27人
学术报告:
2:00pm, July 2: Majority is not Enough: Bitcoin Mining is Vulnerable by EMİN GÜN SİRER
9:00am, July 3: NetKAT: Semantic Foundations for Networks by Dexter Kozen
4:00pm, July 10: The Role of Type Theory in the Use of Proof Assistants for Programming, Software Engineering, and Computing Theory by Bob Constable
4:00pm, July 14 Juris Hartmanis
2:00pm, July 15 - 17 Ken Birman
4:00pm, July 15 Doug James
Normally, the sound in the cartoon are wrong. To produce the sound, people normally just record the sound that they think the true sound is like to be. So Doug James focuses on how to generate the sound better. He presented a lot of projects he has done, like how to generate the sound of water, fire and how to speed them up, how to generate the sound of broken glass which involves plenty of objects, how to model the sound of clothing and so on.
4:00pm, July 16 Eva Tardos
从博弈论发展至今,最有影响力的结果当属纳什均衡,但是纳什均衡仍有很多缺陷,如计算复杂度高、PoA等。Eva Tardos提出了一种新的基于学习的标准代替纳什均衡,在重复博弈中设置regret函数,参与者在regret函数达到threshold时根据best response改变策略,可以证明在所有参与者都采取该策略的情况下social welfare不会比采取纳什均衡的情况差,并且在很多经典案例中都可以得到更好的结果。
11:00am, July 23 Thorsten
Say we have two ranking function for some search engine. When we do experiments to test which function is better, we usually randomly divide people into 2 groups and use some features like #of clicks on the top 1 item to judge. However, this method has two problems. 1. Thorsten’s group has done some experiments and showed that none of the previous metrics truly shows whether a rank function is better than the other. 2. The previous method is silly because it’s just like divide people into 2 groups and ask one group how much they like Coca Cola and ask the other group how much they like Pepsi. What we should do is to ask each person which they like more, Coca Cola or Pepsi. Therefore, they interleaved the results generated by the two functions and showed the combined results to the users. Then they measure something like how many clicks each function receives to compare them. Their experiments show that this method truly shows which function is better. Moreover, based on this interleaving result method, they derived an algorithm to quickly find the best rank function by pairwise comparison only.
11:00am, July 25 Kavita Bala
She works on graphics and introduced 4 main projects. The first two are in rendering. Previously, people need to develop a new model whenever they encounter a new material. Her group developed a method that can systematically handle most kinds of clothing including velvet, silk, wool and so on. It first uses CT scan to obtain the 3d model of the material, then it adjusts some parameters to let the material look like the true material. Finally, it obtains the final rendering result. Since the rendering involves too many details, it’s very slow. So they further proposed a new method to speed it up. The last two projects is on building a database for materials. We now have large datasets for images but not materials. To enhance research on material, they developed the first large-scale database for materials that can extracts the surface of a wood floor, for example, and can obtain its color information, reflectance information, etc. In the fourth project, they build a database which provides pairwise comparison between the true color of the material (disregard of the shading).
2:00pm, July 28 Jon Kleinberg
In social network, people often share images, share blogs and share everything. Some information is widely spread while others are not. When we closely observe the graph of how the message is shared, we can find some structure in the graph that leads to this difference. Therefore, he design a method to predict how well some information will be spread (cascade) given the “share graph” so far. Secondly, we often use the metric - how many friends two people share - to measure how closely related the two people are. However, it’s not often the case. When two people are coworkers, even if they are not good friends, they share a lot of coworkers as friends. So he introduced another metric to measure the closeness of two people. And it showed good performance in predicting a spouse.
11:00am, July 29 Charles F. Van Loan
We have done much study on the matrix. However, nowadays matrix is not enough and people often have to deal with tensors (A tensor is just a multidimensional array and a matrix is a 2-dim tensor). He introduced how people can analyze a tensor and how symmetry in multi-dimension can lead to interesting properties when we unfold a 4-d tensor into a 2-d matrix.
以上学术报告概要,感谢罗璇、杨宽、陈志鹏等同学整理提供
多彩生活: