7. Interpretation Techniques

There are several techniques that we can exercise to learn about the representation a network has learnt during its training:

• Recording activation maps individual or ensemble of kernels

• Lesioning individual or ensemble of kernels

• Probing intermediate layers by linear classifiers