this is for holding javascript data
Mihir Mongia edited abstract.tex
about 8 years ago
Commit id: 8bbccf15e85b06e012236fe7ee0524f9f8c5bbc6
deletions | additions
diff --git a/abstract.tex b/abstract.tex
index 270681d..218fa22 100644
--- a/abstract.tex
+++ b/abstract.tex
...
In the deep learning literature there have been many methods to produce images that correspond to certain classes or specific neurons in the CNN[Zeiler]. There are two main methods in the literature. Deconvolution methods rely on an input image and highlight pixels in an image that activate a neuron of interest. Deconvolution requires the presence of an image. There are other methods that try maximize the class scores or activations of neurons with respect to pixel intensities. However these methods only work for lower level features or more shallow neurons. At higher layers, the neurons represent more abstract concepts such as a dog. Thus an optimal image may have 10 dogs all over the image in different orientations and also have tails and dog ears that don't actually lie on a dog. We propose
a method several potential methods that
does do not rely on an input image and can also create realistic
images of neurons abstract concepts that correspond to
abstract concepts.
Our approach to this problem is to perform a stochastic version of deconvolution that does not rely on a single input image, but instead uses statistics of activations of each layer, produced on a set of images coming from one class. One potential method we could use is using the joint distribution between adjacent layers. To be clear, a certain
set of activations in layer k may fire a neuron in layer k+1. In addition, a certain set of activations neurons "deep" in
layer k-1 may fire the
set of high activations in layer k. CNN.