tag:blogger.com,1999:blog-6141980.post4722209237901029316..comments2024-03-20T12:28:35.004-05:00Comments on Nuit Blanche: Nuit Blanche in Review (March 2017)Igorhttp://www.blogger.com/profile/17474880327699002140noreply@blogger.comBlogger1125tag:blogger.com,1999:blog-6141980.post-60397368465026390782017-04-03T08:30:18.095-05:002017-04-03T08:30:18.095-05:00I made some comments related to compressive sensin...I made some comments related to compressive sensing on this youtube video:<br />https://youtu.be/ceD736_Fknc<br />I guess it would be better to start with a zero weighted random projection net and set the outputs of the reality manifold net to all 1's on real samples during training, rather than starting with a randomly weighted net and learning zeros. The reason is to avoid catastrophic forgetting and better allowing lossy compression of the reality manifold. Well that is just an intuition based on experience with such nets. <br />The idea then is during decompressing you use the reality manifold net to guide the result toward the nearest point on the manifold. <br />I hope that makes some sort of sense!! SeanVNhttps://www.blogger.com/profile/05967727000105480078noreply@blogger.com