- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper analyzes the dynamics of shortcut learning when a finite dataset is reused, in contrast to prior analyses that use fresh samples or population gradients. The setting is a linear predictor trained by full-batch gradient descent with square loss. The data come from a teacher–mixture model with a shortcut direction and a core direction. Using a non-rigorous dynamical mean-field theory (DMFT), the authors track the shortcut and core coefficients and the norm outside their span, which measures sample-specific fitting. The theory predicts when data reuse begins to matter (after about $\sqrt{P/D} steps) and shows that reuse and weight decay can increase shortcut reliance, with simulations matching the theory and image benchmarks giving qualitative support.
Critical strengths
- Comparing the reuse timescale with the shortcut and core learning timescales is interesting and, as far as I know, new in the shortcut-learning literature. In particular, Result 1 and the regimes in Fig. 3 show that the train–test gap and the shortcut-conflicting risk behave very differently depending on how these timescales are ordered.
- The message of each result is stated clearly, and the figures illustrate the different theoretical scenarios well.
- The decomposition into c_sc, c_core and Q_out is a useful tool. It separates "relying on the wrong feature" from "memorizing the training samples", and it shows that the interpolation peak comes entirely from Q_out while the signal coefficients stay finite.
Critical weaknesses
-
Data model. Labels are generated from the full input (Eq. 4), so the shortcut enters the labeling rule itself. This seems unnatural for shortcut learning: in the standard setting (e.g., Sagawa et al., 2020b), the correlation arises from P(s|y), the labeling rule P(y|core) is invariant, and hence a core-only predictor is correct across environments. How does this difference affect the results, and why is the present model a natural choice for studying shortcut learning?
-
Interpretation of Result 2. Since the shortcut and core directions are not orthogonal, the coefficient c_sc is not the same as the component of q along v (e.g., the overlap m), which is what the prediction actually depends on through the shortcut sign (Eq. 9). An increase in c_sc therefore does not necessarily mean that the model relies more on the shortcut. Is the coefficient, rather than the actual magnitude along v, the right measure of shortcut reliance here? Reporting m, or the risk on shortcut-conflicting examples, alongside c_sc would clarify whether the title "can amplify the shortcut" is justified.
-
The experiments do not test the main theoretical results. Fig. 6(a,b) shows that the shortcut coefficient increases when the shortcut agrees with the label more often. This seems trivial, and the paper notes it already happens without data reuse (App. E.3.1). Could you add experiments that test the theory directly? For example, how does the shortcut coefficient change with the weight-decay strength (Result 3)?
-
The role of the DMFT is unclear. For this shallow linear model, the training trajectory already has an exact closed-form solution (Eq. 155), and Sec. 3.2 says Results 2–4 follow "independently of the formal DMFT derivation". So the DMFT seems to be needed only for Result 1, which is an approximate estimate. Could you clarify what the DMFT analysis adds beyond the exact solution in this setting?
Questions for authors
- What is the practical takeaway? Applying DMFT to shortcut learning is novel and interesting, but what problem does it newly solve? My concern is that applying this theory may be the main goal of the paper, rather than addressing an existing problem.
- Result 3: the message that larger weight decay can increase reliance on the shortcut feature is interesting. However, the assumption makes this interpretation difficult, because it means the shortcut direction has larger variance than the core direction. Is this a reasonable assumption in practice?
- Do the predictions hold with pretrained features (e.g., ImageNet-pretrained ResNet embeddings) on Waterbirds, which is the standard setup?
3 comments
Q1. “What is the practical takeaway?”
The practical question is whether the core can be learned before reuse adds appreciably to sample-specific fitting. Fig. 3(b,c) shows this trade-off: with feature geometry held fixed, appreciable reuse precedes core recovery at but follows it at . In the latter case, online learning describes core recovery well; in the former, stopping before reuse would also stop core learning too early. There is also a prediction for a change in the test population. In the fixed-label example at , we kept both fitted predictors fixed and varied only the proportion of shortcut-conflicting test examples. The theory predicts that stronger training agreement loses its average-loss advantage at conflicting examples; the measured crossings were at and at (Fig. 10d).
Q2. Is it reasonable to assume higher variance along the shortcut than along the core?
Yes. Giving the shortcut a larger scale than the core is an established modelling choice. For instance: Puli et al. (2023) scale the shortcut directly; Jain et al. (2024) obtain higher variance along the shortcut by separating the group means; and Hermann et al. (2024) vary the relative feature scales independently of their predictivity.
In Result 3 (), ridge shrinks the higher-variance shortcut direction less, so the shortcut-to-core coefficient ratio increases giving a sufficient condition. App. D.3 shows that adding variance along the core can reverse the effect.
Q3. Do the predictions hold on pretrained Waterbirds features?
Yes. We can show this in the risk reversal tasks. While the submission used random features on Waterbirds (Fig. 6a), in the revision, we trained a linear readout on frozen ImageNet-pretrained ResNet-50 features and tested the same risk ordering as in Result 4 (revised Fig. 16B–D). At , increasing bird–background agreement from to reduced balanced half-squared test loss by at convergence, even though worst-group accuracy fell by percentage points. Stronger agreement made the average readout worse, but it reduced variation in predictions across training sets enough to lower the overall squared loss. At the loss advantage disappeared. These measurements support the risk-reversal mechanism on pretrained features.
Thank you for the detailed responses and the additional experiments. The new calculation with labels fixed by the core and the experiment on the pretrained features, as well as your answer to Q2, addressed my concerns about the data and experimental settings. The reply on the role of DMFT is also clear. The different scenarios in Figure 3 are interesting.
However, some of my concerns remain. For Result 2, the new title is more accurate, but the new analysis shows that a larger shortcut coefficient can come with a smaller response to flipping the shortcut, so it is still unclear what the amplification means for shortcut reliance. Which one should we take as the measure of what hinders generalization? The population risk in Section 2 seems to depend on m rather than c_{sc}, while Results 2 and 3 are stated in terms of the coefficients. Additionally, for Result 3, the ridge sweep on pretrained features didn't show a change in the shortcut-core ratio, so the experiments still don't directly support it. I also still think that the practical takeaway is limited. I therefore keep my score.
We thank the Reviewer for highlighting the comparison between the reuse and feature-learning timescales. In “Invariant labels and shortcut flips”, we tested whether the risk reversal survives when the core alone fixes the label (W1) and whether a larger shortcut coefficient makes predictions more sensitive to a flip (W2). At the Reviewer’s suggestion, we also used “Pretrained Waterbirds features” to test the reversal on ImageNet embeddings (Q3) and the effect of ridge on the fitted coefficients (W3). Both experiments are described in the general reply.
W1. Does the risk reversal survive when changing the shortcut cannot change the label?
Yes. The original model allowed the core and shortcut directions to overlap so we could study coefficient amplification (Result 2). For the rebuttal, we fixed the label from the core, y=sign(z), before drawing the shortcut. For 90% versus 50% training agreement, the new calculation predicts higher average test loss after one update but lower loss at convergence for 1<P/D<1.421 (App. D.5.2, Eq. 199 and following). The mechanism is the same as in Result 4: stronger agreement leaves less training residual to fit in noise directions. At P/D=1.25 and D=512, across 256 paired training datasets, stronger agreement lowered average squared test loss at convergence by 0.255 but raised it by 0.354 on shortcut-conflicting examples (Fig. 10).
W2. “Is the coefficient, rather than the actual magnitude along v, the right measure of shortcut reliance here?”
For sensitivity to a shortcut flip, ∣m∣ is the relevant quantity. Sec. 2.1 distinguishes the overlap m from the coefficient csc. The statement of Result 2 is about csc, but its original title did not make that clear, so we modified it. With overlapping directions, changing the shortcut from −1 to +1 at fixed Gaussian input changes the score by 2m, where m=νcsc+κccore. The additional calculation in App. D.4 shows that, for μ=0 and P<D, the fit can increase csc while reducing ∣m∣. We tested that prediction by measuring both quantities and the loss on conflicting examples on the same training sets. At D=1024 and P/D=0.5 in a moderate-salience setting, the minimum-norm fit increased csc by 0.044 relative to the population fit, yet ∣m∣ fell by 0.154 and conflicting-group squared loss rose by 0.604. We report the flip sensitivity and conflicting-group loss in App. D.4.
W3. “Could you add experiments that test the theory directly?” In particular, how does weight decay change the shortcut coefficient?
Fig. 6(a,b) of the submission varied training agreement on random features, not weight decay. Following the Reviewer’s suggestion, we added a ridge sweep: we held the pretrained Waterbirds features and the 20 training subsets fixed and varied only ridge. At 95% training agreement near interpolation, we could not resolve a change in the shortcut coefficient when we raised the ridge multiplier from 0 to 0.1. At P/D≈1.1 its mean moved from 0.072 to 0.066 (paired difference −0.006; 95% interval [−0.015,0.004]). The interval includes zero. Nor could we resolve an increase in the shortcut-to-core ratio, the quantity in Result 3 (App. E.3.2, Fig. 19C reports its magnitude). What ridge did change was the test loss. Without ridge, the readouts trained at 95% agreement had lower balanced test loss than those trained at 80% because their predictions varied less across training sets. Ridge narrowed that difference in variability enough that the 95% readouts no longer had lower loss (Fig. 19B).
W4. “Could you clarify what the DMFT analysis adds beyond the exact solution in this setting?”
Eq. 155 gives the gradient descent trajectory for one realised training set. DMFT predicts the typical trajectory across draws in the proportional limit, both when the same examples are reused and when each update uses fresh ones. For Result 1, comparing the fixed-dataset and online dynamics identifies an early excess in out-of-plane fitting proportional to t(t−1)/(P/D). Fig. 3 compares its onset with core recovery: the same feature geometry admits both early and late reuse. Results 2–4 have separate least-squares arguments; they do not require the formal DMFT derivation.