Here, for each filter, we collect input images that cause the filter to respond the most strongly. One can infer what each filter does by the images that are collected. For example, one may pick up on green while another may pick up on red, circular patterns.
tarabyte
As we increase the layer number, the patterns we can detect get more complex!
apk
Layer 1 collects responses to simple gradients, then layer 2 responds to combinations of layer 1 responses (combining them using the weights that were established in the training stage), and then layer 3 does a similar thing with layer 2 responses.
kayvonf
Images taken from the following paper:
Visualizing and Understanding Convolutional Networks. Zeiler and Fergus. ECCV 2014
What exactly does it mean to respond most strongly to? What is being measured to determine what is strong vs. weak?
rsvaidya
@yes I think when they say respond most strongly, it means after the filter is applied the value of the pixel in that area is higher compared to images where that patter is not found.
Here, for each filter, we collect input images that cause the filter to respond the most strongly. One can infer what each filter does by the images that are collected. For example, one may pick up on green while another may pick up on red, circular patterns.
As we increase the layer number, the patterns we can detect get more complex!
Layer 1 collects responses to simple gradients, then layer 2 responds to combinations of layer 1 responses (combining them using the weights that were established in the training stage), and then layer 3 does a similar thing with layer 2 responses.
Images taken from the following paper:
Visualizing and Understanding Convolutional Networks. Zeiler and Fergus. ECCV 2014
https://arxiv.org/abs/1311.2901
What exactly does it mean to respond most strongly to? What is being measured to determine what is strong vs. weak?
@yes I think when they say respond most strongly, it means after the filter is applied the value of the pixel in that area is higher compared to images where that patter is not found.