From 0fe5713f832be0967f9a7511b0f337e64ecd1ab4 Mon Sep 17 00:00:00 2001 From: mayukhdeb Date: Sat, 24 Feb 2024 12:19:06 +0530 Subject: [PATCH] new post wip --- .../post/2024-02-23-magnitude-constrained-featurevis.md | 8 ++++++++ 1 file changed, 8 insertions(+) create mode 100644 content/post/2024-02-23-magnitude-constrained-featurevis.md diff --git a/content/post/2024-02-23-magnitude-constrained-featurevis.md b/content/post/2024-02-23-magnitude-constrained-featurevis.md new file mode 100644 index 0000000..ed8ebcc --- /dev/null +++ b/content/post/2024-02-23-magnitude-constrained-featurevis.md @@ -0,0 +1,8 @@ +The primary drawback of feature visualization has been it's inability generate interpretable features in deeper networks. In my own experience, I've seen that feature vis basically stops working once we go past the 3rd resnet block in a resnet18. + +This paper somehow fixes this issue by optimizing the images in the phase spectrum while keeping the magnitude constant. + +**What is the phase spectrum?** + +We know that the fourier transform provides the frequency domain information about the signal. The phase spectrum provides the individual phases of each individual frequency component in the frequency domain representation. To gain a better understanding, let's visualize it in a few plots. +