Optimizing stimuli to Test models of perception

Creating metamers from a trained model:

• Start with a target image …eg a dog.
• inject noise look at teh response of a given layer… It will be differnet than the dog. Change the input noize (through back prop ?) until the input matches the response in the latent space Q: Is there a unique way to do this? eg: isnt there a many-to-one between input data and latent space. Q: How specific are metamer to the network?: eg:

Dont the adersarial examples (trivial amount of specific noise changes a class assignment) show that the NN vision in nothing like humans ? Claim the brian doenst suffer from adverserial examples b/c we train in a nosy envirioment. JA: Suggests that this is a better way of training to remove the adverserial attacks (as opposed to explicit adversiral training..)