Totally nice to have, to improve interpretability.
[ ] Visualize convolution filters
[ ] Visualize feature maps
Try out
[ ] GradCam
[ ] Other visualization techniques?
One would expect the network to learn very structured features (like harmonics or circles ?) when training on deadleaves
But may experience something totally different when training on natural data?
Totally nice to have, to improve interpretability.
Try out
One would expect the network to learn very structured features (like harmonics or circles ?) when training on deadleaves But may experience something totally different when training on natural data?