- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper provides an algorithm to certify whether a conditional location-scale mixture can generate distinct distributions under two different conditioning inputs and gives a quantitative lower bound on the distance between them if passing certificates. The experiments give some numerical results to give some evidence for the proposal algorithm.
Critical strengths
- Developing OOD generalization theory for conditional generative models is widely recognized as a challenging and important problem, so the intended problem is an interesting research question.
Critical weaknesses
- This paper’s main weakness is its poor presentation, characterized by a confusing logical structure and a clear lack of adherence to academic writing conventions. This gives the impression that the manuscript has not been carefully polished.
- This paper does not demonstrate that certifying such a lower bound is meaningful for practical conditional generative models.
Questions for authors
- What is the most important theoretical result presented in the main text? It is difficult to understand the results.
- It is unclear what the experiments are intended to demonstrate. I do not see how they support the proposed algorithm or the theoretical results. You should explain the meaning of experiments in the beginning of this section.
- The results are derived for conditional location-scale mixtures, can you provide certifying algorithm (at least some insights) for general conditional generative models?