Questions for "OpenGAN: Open-Set Recognition via Open Data Generation"
-----------------------------------------------------------------------------------------------

Please send your answers to: galessos@cs.uni-freiburg.de by 13:30 on 07.02.2022

Question 1. (2 sentences) Previous works have found GAN discriminators to be ineffective for open-set recognition. Why? How does OpenGAN circumvent the issue?

Question 2. (2 sentences) Consider the training objective in equation 2. What role does the second term of the loss play? In general, what is the effect of setting lambda_o=0?

Question 3. (2-3 sentences) What is, in your opinion, the main strength of the paper's evaluation section? Can you identify any weakness?