What this demonstrates
The output at each location is a dot product between a local image patch and a filter. With hand-crafted filters you can see classical image operations. With learning, gradient descent updates the filter so that plus-sign regions become bright.
Images
Input image. Size:
Target map. Size:
Filter response. Size:
Thresholded output.
Threshold output
0.75
Shared grayscale colour bar for all image panels. Range: 0 to 1
Kernel and loss
Current 5×5 kernel with values shown in each cell.
Training loss history.
Current loss
-
Peak response
-
Kernel L2 norm
-
Mode
Manual