I found the idea of "designing" adversarial examples for a pretrained model very interesting, especially when noise was treated as a trainable object.
The experiments with different values of epsilon in conjunction with the very last figure in the report were interesting and insightful.
Areas of Improvement
I don't see the point of comparing model probabilities to a uniform distribution (cell 406). It might have been useful in exploratory analyses, but does not tell much to the reader.
It would be interesting to add a few dense layers on top of the pretrained model, set those layers to be trainable, and then train such a model on noisy images. I'm not sure, but perhaps with enough examples the model can learn to ignore the noise and see the object behind it.
Feedback on Biweekly Report 4
By Behzad Vahedi
Superpowers
Areas of Improvement
Price
A