NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:9327
Title:On Relating Explanations and Adversarial Examples

Reviewer 1

The work is amazingly well-written and easy to follow. The work does a pretty good job of introducing the required logic background for understanding the rest of the work (I was not introduced to the notions and terminology before and I found it very easy to follow). As a researcher in both fields, I find this new direction very much and appreciate the work's significance. However, from a practitioner's point of view, the main question is the applicability of the introduced algorithm. The algorithm requires computations up to the order of the feature space size (which is intractably large even for the simplest problems). I am leaning towards an acceptance score due to its novelty but the actual contribution and its usefulness in practice should become more clear.

Reviewer 2

This paper address an intuition that has been present in the literature for some time, but has not been formalized or published that i am aware of. Many papers hint at the duality between adversarial examples and explanations, but by formalizing these notions and proving the duality between them, the authors make an important contribution to the literature, the paper is quite dense, and spends most of its time on theoretical proofs and justifications for the relationship between adversarial examples. It may find a wider readership if it allocates some more space to introducing the concepts it employs, even such simple things as "subset-minimal" may not be widely known in the adversarial community. The final section focusing on the experimental result making use of a subset of a binarized version of MNIST is compelling, and makes the theoretical work significantly easier to grasp. The running example of the restaurant running problem, also helps illustrate the theorems presented, but the paragraph describing the specifics of the problem is probably not needed. Overall i think this paper is an important contribution to the research literature, but could be made more approachable and subsequently reach a broader audience with some light rewriting with a focus on making more accessible.

Reviewer 3

Overall Comments First I should state upfront that my expertise is not in first order logic, so it is somewhat difficult to assess this paper. I am familiar with the literature on adversarial examples and explanations though. My main high level point in this work is that I am missing a so what, i.e., what is the significance of demonstrating the duality between explanations and adversarial examples. In addition, there has been intense empirical work showing that explanations can be used to craft adversarial examples and adversarial examples might be good instances for producing explanations. Originality The main point of this paper has been demonstrated empirically in prior work. The authors present a set of theorems based on formal logic demonstrating a duality between adversarial examples and explanations. In terms of the theorems, I am not familiar with the formal logic literature , though in looking at citations [21-23], it does seem that these theorems are new. Clarity The work is reasonably well written and free of typos. Several of the key formal logic terms are also defined and clarified. Theorem 1 was clearly stated and the proof clarified in the text. A proof sketch was also provided for theorem 2. Significance In looking at the prior work, citations [21-23], it seems the theorems here are new. In general, there have been previous connections made between explanations and adversarial examples:, and, however these were not formal. In fairness, it is hard for me to assess the significance of this work since it seems like the key insight is bringing the FOL point of view to clarify the relationship between adversarial examples and explanations. Some Issues - The MNIST Example. I sense this is probably the wrong example to use to show the power of your analysis. It seems like MNIST is high-dimensional for the logic based models or decision set type framework. It would've been more powerful for me, if the paper had shown their results on lower dimensional dataset with enough categorical variables to show the power of this work. The current MNIST example feels toyish. - Can the authors further clarify the implications of the duality that they motivate in this work? UPDATE I have read the author rebuttal and feel the authors provided justification for their approach and why this work is important. I support that this work be accepted.