- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This work studies when a gain in predictive log score supports a claim of additional information when adding new input, a common move in machine learning to improve model performance. But improvements from adding new input may just mean the original model was a poor approximation. The work provides a mathematical proof that we need a guaranteed upper bound on how wrong the original model is (a “witness”) in order to claim additional information
From there they show that the usual diagnostics, such as reverse-KL error, can look promising even when the gain reflects no new information. For categorical predictions, they derive an exact reverse-to-forward KL bound requiring only the smallest baseline probability and the solution of one scalar equation. The audit combines this error bound with allowances for witness failures and sampling uncertainty to give a lower confidence bound on conditional information. Controlled experiments show that it avoids certifying improvements from a deterministic representation while detecting an added informative signal. On WDBC, Digits, and Wine, it produces positive certificates in six of nine analyses.
Critical strengths
- The main theorems and proofs in the work appear correct, as far as I was able to verify. The paper makes a clear and formal distinction between a model making better predictions and an input adding information, and what better predictions establishes about the new input. This is a point often overlooked in empirical research, such as when interpreting feature or representation gains. The separation example makes the issue of standard checks concrete.
- The exact categorical bound is a useful theoretical contribution. It replaces a search over class splits with solving a single scalar equation, and Supplementary Table A4 shows that the generic divergence bound is approximately 1.5–2.4 times larger in the tested settings.
Critical weaknesses
- The practical applicability of the results seem to be limited, as it is unclear how often a useful witness can actually be obtained. The method needs a justified upper bound on the baseline’s error, which is also the hardest requirement to satisfy. Every experiment gets it from something convenient: a known teacher model (learned representation), an enumerable posterior (the Bayesian study), or a context with only two groups, so the probabilities can be estimated directly (natural labels). These examples help explain the method, but leave open how to obtain a sufficiently tight witness for richer inputs when the true outcome distribution is unknown.
- The experiments do not clearly establish the downstream benefits of the main theoretical results. The successful natural-label experiments construct their witnesses directly from outcome counts, while the learned-representation experiment uses known outcome probabilities. I don’t think these results necessarily show whether the new categorical bound makes otherwise inconclusive information claims certifiable. A comparison holding the predictors, approximation budget, and audit data fixed would help establish how much the sharper bound changes the final conclusion.
Questions for authors
- Beyond the settings demonstrated in the paper, where else do you expect a sufficiently tight witness to be obtainable in practice? For example, what additional assumptions, validation data, or error guarantees would you need to construct one for the time-series application?
- How much difference does the categorical bound make to the final audit result? With the predictors, approximation budget, and audit data held fixed, how much does it improve the information lower bound over a generic divergence bound, and how often does it turn an inconclusive result into a positive certificate?