tag:blogger.com,1999:blog-7345806147365425073.post2232594127522453494..comments2020-05-18T07:24:28.399+02:00Comments on Peekaboo: [CVML] Florent Perronnin on explicit feature mapsAndreas Muellerhttp://www.blogger.com/profile/10177962095394942563noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-7345806147365425073.post-65836490771330643832011-08-18T16:16:56.634+02:002011-08-18T16:16:56.634+02:00Well, he also presented the Fisher kernel, which i...Well, he also presented the Fisher kernel, which is a more ad-hoc embedding that works well in practice.<br />Perronnin's research focus seems to shift in this direction so maybe he'll agree with you there.<br /><br />On the other hand, being able to model arbitrary decision boundaries as the rbf kernel can is kind of a nice feature ;)Andreas Muellerhttps://www.blogger.com/profile/10177962095394942563noreply@blogger.comtag:blogger.com,1999:blog-7345806147365425073.post-33260533318001172052011-08-18T16:12:37.057+02:002011-08-18T16:12:37.057+02:00"So is approximating non-linear kernels the w..."So is approximating non-linear kernels the way to go? It seems some researchers think it is...<br />But wasn't the point of using the kernels that they are easy to compute? So maybe we should instead look for clever embeddings that work well in practice and stop caring about kernels at all ?<br />We'll see."<br /><br />This is exactly the point I was thinking about after his lecture. Why not look directly for ways to efficiently extract features from a compressed representation?Urinoreply@blogger.com