search

tdeboissiere / DeepLearningImplementations

Implementation of recent Deep Learning papers
MIT License
1.81k stars 650 forks source link

Deconv color meaning #16

Closed DanlanChen closed 7 years ago

DanlanChen commented 7 years ago

Hello, I am wondering if the colors (back ground color in each deconv feat map represent some meaning, like which feature map in the original image has more action response in the neuron, like which part of one image does the network reponses most?

tdeboissiere commented 7 years ago

The Deconvolution here consists of identifying which image maximally activates a specific neuron and then trying to reconstruct the part of the image that contributed to this activation.

I don't know whether the background color is an artifact or if it has a clear interpretation. In Zeiler's paper, the background is mostly gray...

You could try to plug the reconstructed image into the network again, change its background color and check whether it still activates the neuron strongly.