tag:blogger.com,1999:blog-2293611840857369720.post231856070345338416..comments2024-10-30T14:44:00.251+03:00Comments on Computer Blindness: Max-product optimization methods and softwareAnonymoushttp://www.blogger.com/profile/01422787855469629259noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-2293611840857369720.post-8680450211430224322011-09-09T03:11:58.328+04:002011-09-09T03:11:58.328+04:00Best explanation of SDP for integer programming is...Best explanation of SDP for integer programming is Chapter 6 of http://www.designofapproxalgs.com/book.pdf<br /><br />Maybe you would need a specialized SDP solver to be practical. For instance, Weinberger in "Distance Metric Learning for Large Margin Nearest Neighbor Classification" developed their own SDP optimizer. Also, I was recommended the following algorithm for cheaply solving the SDP approximately -- http://www.cs.berkeley.edu/~satishr/cs294/scribe/2-17.pdfYaroslav Bulatovhttps://www.blogger.com/profile/06139256691290554110noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-69613964605535196342011-08-30T02:36:11.176+04:002011-08-30T02:36:11.176+04:00I have not seen anything like this so far.
From a ...I have not seen anything like this so far.<br />From a quick look at the Wikipedia article it is not clear to me how to use it.<br />LP relaxation is not generally used directly since it is slow, but the methods like TRW and submodular decomposition optimize the same function.<br />So how do you think, is it enough to formulate the problem in terms of SDP, or we also need to invent optimization techniques?Anonymoushttps://www.blogger.com/profile/01422787855469629259noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-7189218037976953552011-08-29T13:24:02.257+04:002011-08-29T13:24:02.257+04:00Has anyone in vision tried Semidefinite Programmin...Has anyone in vision tried Semidefinite Programming for finding lowest energy labeling? SDP relaxation is more powerful than LP relaxation, and it's straightforward to setup given access to SDP solver like CVXOPT -- http://mathematica-bits.blogspot.com/2011/03/semidefinite-programming-in-mathematica.htmlYaroslav Bulatovhttps://www.blogger.com/profile/06139256691290554110noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-1129178140720700682010-06-02T02:42:51.736+04:002010-06-02T02:42:51.736+04:00You are right, in theory the potential functions c...You are right, in theory the potential functions correspond to the factors, but in practice the only constraint is non-negativity (before taking logarithms; exp(φ) > 0 in my notation), so they can be even greater than one. Probabilistic reasoning is desirable, but not always practical.<br />You are also right about the likelihood. That is not likelihood in mathematical sense. Since I am trying just to give the intuition, I replaced it with the neutral "possibility".Anonymoushttps://www.blogger.com/profile/01422787855469629259noreply@blogger.comtag:blogger.com,1999:blog-2293611840857369720.post-16685770133295619502010-06-01T00:30:08.882+04:002010-06-01T00:30:08.882+04:00>> So-called potential functions φ (unary in...>> So-called potential functions φ (unary in the first term and pairwise in the second one) define likelihood of the assignment given class labels to the nodes and edges correspondingly. They could be conditioned by the data.<br /><br />We may have some misunderstanding here. As I always thought, potential functions correspond to the factors of the whole probability density function, not to some "likelihood". Then, likelihood function appears when data is conditioned by (on?) something, not something by data.hr0nixhttp://sexdrugsandappliedscience.comnoreply@blogger.com