- Rating
- 1
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper argues that analyses like pruning scores or edit-acceptance checks should give the same answer on two networks that compute exactly the same function. It provides a formal framework and then it tests this by applying function-preserving changes to the weights (permutations, rescaling) and rerunning the analyses.
Critical strengths
S1. The motivating framework is sensible. Asking which claims based on previous techniques (pruning scores, importance rankings, etc.) are invariant to a function-preserving reparameterization is an interesting question in itself, and the paper's framing is not without interest. For instance, a rescaling that preserves the function can still change which layers get pruned (Fig. 2) or whether an edit is accepted (Table 1), while the final quality stays almost the same. These results should have been the main ones of the paper.
Critical weaknesses
W1. What the framework presented serves for is unclear. The full fibers and quotients are never computed (to be fair, acknowledged at l.485), as it only uses some specific transformations and permutations. Moreover, the results (Prop 1 and 2) are elementary and do not add much. Finally, it is not clear how it helps with knowing which claims should hold at the function level. The conclusion does not help much with this -- it just says again what the framework is.
W2. The experiments and the guidelines in 4.6 do not add much insight. Figure 1(a) is a sanity check not for experiments, and the NLL differences (the most interesting results) are small. Importantly, the guidelines in 4.6 are generic: there is no way to know how they result from the framework and the experimental insights. They should at the very least be linked to some quantitative results.
W3. Much of the information important for the paper is hidden in the appendix, so a reading of the first 9 pages is not enough. For instance, "selector" and "intervention gap" are used in Section 4.2 without being defined in the main text.
Questions for authors
Q1. How can one find a certified transformation on a new architecture without listing all possible transformations?
Q2. What is the decision procedure to know whether a claim should hold for a function only or for a given representation? (take the example of explaining the architecture itself which depends on the representation)