tag:blogger.com,1999:blog-19803222.post3373191565627655987..comments2018-02-16T22:28:23.905-07:00Comments on natural language processing blog: Column squishing for multiclass updateshalhttp://www.blogger.com/profile/02162908373916390369noreply@blogger.comBlogger5125tag:blogger.com,1999:blog-19803222.post-35034960049196536912017-08-20T08:17:31.147-06:002017-08-20T08:17:31.147-06:00This comment has been removed by a blog administrator.hamid alidoostihttps://www.blogger.com/profile/07646712673915758900noreply@blogger.comtag:blogger.com,1999:blog-19803222.post-78662290845443015732017-08-19T08:34:36.076-06:002017-08-19T08:34:36.076-06:00Do you mean Figure 2 in http://jmlr.csail.mit.edu/...Do you mean Figure 2 in http://jmlr.csail.mit.edu/papers/volume7/crammer06a/crammer06a.pdf ? This doesn't seem to be an exact solution to the problem I wrote above unless I'm missing something. It only updates two classes, which is going to be insufficient in general. Probably I'm looking at the wrong paper tho because Figure 3 in that paper is a graph :/. Can you point me in the right direction?<br /><br />I'm not sure why you refer to this as an approximation. Which step is approximated?halhttps://www.blogger.com/profile/02162908373916390369noreply@blogger.comtag:blogger.com,1999:blog-19803222.post-31044033599769556652017-08-18T17:18:36.538-06:002017-08-18T17:18:36.538-06:00You can solve the Lagrange / Fenchel dual of your ...You can solve the Lagrange / Fenchel dual of your setting exactly. It was re-derived multiple times starting with Kesler. Crammer and myself gave an explicit algorithm for both the separable and the non-separable case. We provided a d time exact algorithm as well as a fixed point algorithm (figures 2 & 3). Your approximation is very nice thought I suspect not nearly as fast when the number of classes is in the thousands.Unknownhttps://www.blogger.com/profile/07847770619575725635noreply@blogger.comtag:blogger.com,1999:blog-19803222.post-19912090471353385752017-08-16T06:42:34.368-06:002017-08-16T06:42:34.368-06:00Also check out these two papers (h/t Mathieu Blond...Also check out these two papers (h/t Mathieu Blondelâ€Ź):<br /><br />http://epubs.siam.org/doi/abs/10.1137/1.9781611972801.27<br /><br />http://mblondel.org/publications/mblondel-icpr2014.pdfhalhttps://www.blogger.com/profile/02162908373916390369noreply@blogger.comtag:blogger.com,1999:blog-19803222.post-19249373372056563902017-08-15T11:41:02.467-06:002017-08-15T11:41:02.467-06:00The fast algorithm is over 100x faster than a nume...The fast algorithm is over 100x faster than a numerical solution (and produces the same result).<br /><br />https://gist.github.com/timvieira/4a4e7e700c34c04160b93aa03a14861cTim Vieirahttp://timvieira.github.io/blognoreply@blogger.com