tag:blogger.com,1999:blog-6950833531562942289.post3284554322984297797..comments2024-03-25T03:36:48.099-07:00Comments on C0DE517E: A programmer's sightseeing tour: Machine Learning and Deep Neural Networks (part 1)DEADC0DEhttp://www.blogger.com/profile/01477408942876127202noreply@blogger.comBlogger3125tag:blogger.com,1999:blog-6950833531562942289.post-83542412537349295142017-03-29T22:35:07.160-07:002017-03-29T22:35:07.160-07:00Unfortunately, not really. Of course there are cer...Unfortunately, not really. Of course there are certain rules, like certain activation functions are better for classification rather than regression, certain network characteristics are desirable when we know our problem has some invariants and so on and so forth... But overall that's the catch of deep NN, they replace searching for good features with searching for good NN architectures and training schedules. Best bet is to study NNs that worked in the past for a similar problem and adapt them. If you look at DNN papers often all they do is to describe the architecture, because then again that's the part that requires human invention.DEADC0DEhttps://www.blogger.com/profile/01477408942876127202noreply@blogger.comtag:blogger.com,1999:blog-6950833531562942289.post-90338057008378163002017-03-27T14:46:41.197-07:002017-03-27T14:46:41.197-07:00Typo: interesting read ...Typo: interesting read ...Ashhttps://www.blogger.com/profile/12625252236983121298noreply@blogger.comtag:blogger.com,1999:blog-6950833531562942289.post-77283145740763600212017-03-27T14:45:36.115-07:002017-03-27T14:45:36.115-07:00Interested read, thanks for sharing. What always h...Interested read, thanks for sharing. What always has seemed like black art to me is how this networks are designed. Surely you can't just randomly append layers and expect to get a good result. Do you have any insights on how to go about designing a network that achieves a particular goal. Say a segnet - how do you design the network to segment an image?Ashhttps://www.blogger.com/profile/12625252236983121298noreply@blogger.com