1. Visualize patches that maximally activate neurons
找到图片的区域能最大激活某个神经元
2. Visualize the weights
3. Visualize the representation space (e.g. with t-SNE)
取出某一层的output feature,用t-SNE 投影进行visualization
4. Occlusion experiments
5. Deconv approaches (single backward pass)
compute the gradient of an arbitrary neuron in the network w.r.t. the image
用guided的方法进行backprop可以得到更好的结果
6. Optimization over image approaches (optimization)
find an image that maximizes some class score(用梯度更新图片的值,这与普通的backprop更新weight不同)
7. Given a CNN code, reconstruct the original image