- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes RPS-TrReg, a representer point method that explains the predictions of a trained classifier in terms of its training samples. The authors argue that RPS-l2 explains an l2-regularized last layer rather than the trained model, and that RPS-LJE and influence functions need the inverse Hessian. RPS-TrReg freezes the feature extractor and re-trains the last layer on the training loss plus a penalty that keeps it close to the original last layer. In the derivation this re-training starts from zero. By unrolling the gradient descent steps and using the stationarity condition at the last step, the authors then write the original last layer as a linear combination of the training features. Each training point gets a global importance, computed from its loss gradients along the fine-tuning path. Its local importance is the inner product between its features and those of the test point. The method is evaluated on a binary horse vs. car subset of CIFAR-10 with a ResNet-20 and on IMDB sentiment classification with a Bi-LSTM, against influence functions, RPS-l2, RPS-LJE, TracIn and NTK-Trac. In the main experiment, 30% of the training labels are flipped, training points are inspected in order of their global importance and the model is retrained after each 5% is corrected. RPS-TrReg finds more flipped labels than the baselines on CIFAR-10 and slightly more on IMDB, and gives the highest accuracy after correction at most inspection levels. The authors also report the L1 distance between the reconstructed and the original last-layer weights, where RPS-TrReg is lowest. In runtime, RPS-TrReg is slightly slower than RPS-l2 and NTK-Trac but faster than influence functions and RPS-LJE.
Critical strengths
The main strength is that the paper addresses a real problem, since RPS-l2 explains an l2-regularized last layer rather than the model that was actually trained. Anchoring the penalty to the original weights instead of to zero is a simple fix, and the derivation is short and easy to follow. The algebra in Eqs. 9-14 is correct when the last layer starts at zero. The anchoring idea could also be useful for other representer methods. Also the cost is low. The method only needs per-sample gradients of the last layer along one fine-tuning run and no Hessian. In Table 1 it is between 3% and 21% slower than RPS-l2 and NTK-Trac and between about 2.4 and 8 times faster than influence functions and RPS-LJE, so it would be cheap to use in practice.
Critical weaknesses
The first weakness is that the experiments do not test whether RPS-TrReg explains the original model better. The only direct measure of fidelity is Figure 2, the L1 distance between the reconstructed and the original last-layer weights. From Eqs. 7, 10 and 13, with a zero start and full-batch steps, this difference is exactly the gradient of the fine-tuning objective at the last step divided by lambda k^T. For RPS-TrReg the figure therefore measures how well the fine-tuning converged, and for RPS-l2 it mostly measures how far its weights drifted from the original ones. Further, with many more training points than features, infinitely many weightings reconstruct the original weights exactly, so a good reconstruction does not show that the attributions are correct. The data debugging experiment ranks training points by the global importance alone. This score does not depend on the test point and is built only from each point's own loss gradients, so it may largely track the training loss. Koh & Liang (2017) use the training loss as a baseline for this task, but the paper does not include it. Since fidelity is the central claim of the paper, a counterfactual evaluation, such as retraining without the top-ranked points or the linear datamodeling score of TRAK, is needed. The paper also does not report the learning rate, number of steps or batch size, and gives no error bars in Figure 1, where the IMDB differences are small.
Also I have a concern about the limited novelty and the small setting. The discounted gradient sum in Eq. 10 has the same form as the fast approximation in HyDRA, which keeps the L2 regularization and drops the Hessian. It is also closely related to the tracking representers of Tsai et al. (2023a), which they evaluate with a last-layer kernel. The new parts are the anchor to the original weights and the use of the stationarity condition to solve for them. However, the paper does not show which of these matters, for example by dropping the anchor or the terminal term separately. MAGIC is described as global and orthogonal to this work, but it attributes individual predictions and I think it is the closest strong baseline. The paper is motivated by the cost of Hessian-based methods on large networks, but all runtimes in Table 1 are below 1.6 seconds. The speed-up over influence functions and RPS-LJE is also smaller on IMDB (about 4 and 2.4 times) than on CIFAR-10 (about 6 and 8 times), which does not support the claim that the advantage grows with dataset size.
Questions for authors
- How was the last layer initialized in the experiments, and was the fine-tuning full-batch as in Eq. 8, even though the text says SGD?
- Are training points in Figure 1 ranked by the signed or the absolute global importance, and how does RPS-TrReg compare with ranking by training loss?
- Is the "total influence" used for influence functions and TracIn in Figure 1 the self-influence that Koh & Liang (2017) and Pruthi et al. (2020) use for this task? TracIn ranks below random selection on CIFAR-10, which would be unexpected with self-influence.
6 comments
Weakness 2.1 & Weakness 2.3: Difference to HyDRA and Consider MAGIC as a Baseline. Unlike RPS-TrReg, the regularization in HyDRA is not designed to anchor reconstructed weights to the original target parameters , nor does HyDRA perform explicit weight reconstruction. Instead, HyDRA includes regularization in its training objective, this term contributes to the gradient trajectory and appears in the assumptions used to bound the approximation error when replacing the inverse-Hessian term with a Hessian-free hypergradient approximation. In contrast, RPS-TrReg uses trajectory regularization specifically to keep the optimization path anchored to the target model's parameters without needing to approximate the Hessian at all.
Following your suggestion, we added both HyDRA and MAGIC as baselines to our original data debugging benchmarks (Bi-LSTM on IMDB and ResNet-20 on CIFAR-10) and conducted additional experiments using GPT-2 on SST-2 and ViT-B/16 on MNIST (binary classification of even vs. odd digits). For a fair comparison, MAGIC and HyDRA use the same backtracking line search protocol as RPS-TrReg. As shown in the tables below, which report the percentage (%) of mislabeled data identified at 10%, 20%, and 30% data inspection, while MAGIC achieves the highest score on CIFAR-10, RPS-TrReg consistently outperforms both baselines across the remaining three benchmarks. We will include these comparisons and the discussion of HyDRA in the revised manuscript.
Bi-LSTM on IMDB Results:
| Method | 10% Checked | 20% Checked | 30% Checked |
|---|---|---|---|
| HyDRA | 13.29 ± 1.01 | 25.56 ± 1.44 | 37.05 ± 1.69 |
| MAGIC | 13.91 ± 0.80 | 26.19 ± 1.09 | 37.46 ± 1.40 |
| RPS-TrReg | 14.88 ± 0.72 | 27.46 ± 0.88 | 38.53 ± 0.95 |
ResNet-20 on CIFAR-10 Results:
| Method | 10% Checked | 20% Checked | 30% Checked |
|---|---|---|---|
| HyDRA | 19.33 ± 2.30 | 32.65 ± 2.75 | 43.09 ± 3.74 |
| MAGIC | 25.00 ± 2.30 | 40.08 ± 4.70 | 52.50 ± 6.70 |
| RPS-TrReg | 21.89 ± 3.06 | 35.80 ± 4.05 | 47.61 ± 4.09 |
ViT-B/16 on MNIST Results:
| Method | 10% Checked | 20% Checked | 30% Checked |
|---|---|---|---|
| HyDRA | 32.94 ± 0.20 | 56.06 ± 0.98 | 65.06 ± 1.19 |
| MAGIC | 33.13 ± 0.06 | 65.30 ± 0.35 | 88.21 ± 1.97 |
| RPS-TrReg | 33.20 ± 0.03 | 66.12 ± 0.10 | 94.32 ± 0.39 |
GPT-2 on SST-2 Results:
| Method | 10% Checked | 20% Checked | 30% Checked |
|---|---|---|---|
| HyDRA | 17.18 ± 1.51 | 29.03 ± 2.23 | 37.32 ± 2.84 |
| MAGIC | 23.26 ± 1.04 | 41.02 ± 1.46 | 55.40 ± 1.61 |
| RPS-TrReg | 30.19 ± 0.10 | 54.90 ± 0.18 | 72.55 ± 0.32 |
Weakness 2.2: Experiment of Dropping the Anchor. Thank you for this insightful question. In our experiments, our implementation of the NTK-Trac baseline is mathematically equivalent to RPS-TrReg without the proximity anchor term (). As shown in Fig. 1 and Fig. 2, removing this anchor term leads to a noticeable drop in data debugging performance on both CIFAR-10 and IMDB, as well as a substantially larger weight alignment gap, because the trajectory is no longer bound to and fails to faithfully approximate the original model weights. The following tables present the data debugging comparison between our method with and without the anchor term included:
Bi-LSTM on IMDB Results:
| Method | 10% Checked | 20% Checked | 30% Checked |
|---|---|---|---|
| NTK-Trac (dropping anchor) | 14.19 ± 0.96 | 25.76 ± 0.84 | 36.21 ± 0.82 |
| RPS-TrReg | 14.88 ± 0.72 | 27.46 ± 0.88 | 38.53 ± 0.95 |
ResNet-20 on CIFAR-10 Results:
| Method | 10% Checked | 20% Checked | 30% Checked |
|---|---|---|---|
| NTK-Trac (dropping anchor) | 18.62 ± 2.22 | 33.12 ± 3.39 | 45.77 ± 4.49 |
| RPS-TrReg | 21.89 ± 3.06 | 35.80 ± 4.05 | 47.61 ± 4.09 |
Weakness 2.4: Comprasion to the cost of Hessian-based methods. We thank the reviewer for pointing this out. We agree that the runtimes in Tab.1 do not show that our speed advantage grows with dataset size. These experiments use binary classification with a frozen feature extractor, so the influence-based baselines compute only the final-layer Hessian: . Here, n is the number of training examples, is example i's feature vector, is its predicted positive-class probability, I is the identity matrix, and is numerical damping, d is the feature dimension. This is the exact Hessian for the binary logistic loss with respect to the frozen linear head. In our experiments it is .
For a -class head, a dense Hessian over all head parameters expands to , making exact inversion substantially more expensive. Relative to the binary-head case, its storage scales by and dense inversion by (about and , respectively, for ). Since we have not measured those multi-class settings, we will qualify our scalability claim in Section 4.4 to state that Tab. 1 demonstrates efficiency for frozen-feature binary classification, and we will remove the claim that the speed advantage scales with dataset size.
Question 3: Total influence definition and TracIn performance below random. Thank you for pointing this out. In Fig.1, the Influence Function baseline ranks examples by the magnitude of their self-influence, , where is the per-example gradient and H is the Hessian. The TracIn score is the self-influence specialization , accumulated over 30 updates.
The possible explaination of TracIn ranks below random is that early erratic gradients caused by fitting contradictory flipped labels disproportionately inflate scores for clean but boundary examples, while late-stage memorization drives down the gradients of already-fitted mislabeled samples. Unlike Influence Functions (which normalize curvature via ) or RPS-TrReg (which anchors the trajectory to the final model state ), unregularized squared gradient norms become dominated by optimization noise.
Weakness 1.1 & Question 2: Lack of Training Loss Experiments. We thank you for this crucial observation. In Fig. 1, we rank training samples by global importance. Following the reviewer’s recommendation, we conducted leave-one-out (LOO) experiments. For each of ten randomly selected test samples, we ranked the training samples by signed importance. For RPS methods, signed importance is defined as global importance×local importance. We removed each training sample in turn, retrained the model, and measured the change in the test sample’s loss.
For each test sample, we calculated the Pearson correlation r between the method’s signed importance scores for the training samples and the observed test-loss changes after removing each training sample and retraining. We report the mean and standard deviation of these correlations across the ten test samples.
As shown in the tables below, RPS-TrReg achieves the highest mean Pearson correlation on both datasets and outperforms Influence Functions without requiring Hessian computation. MAGIC and HyDRA also perform strongly. Their trajectory-based attributions capture how training examples affect parameter updates across optimization steps, while Hessian-related information captures local curvature. Together, these may help explain their stronger correlations with the observed loss changes.
Bi-LSTM on IMDB Results:
ResNet-20 on CIFAR-10 Results:
Weakness 1.2 & Question 1: Optimization details & full-batch vs. SGD. Thank you for raising this question and helping us improve the clarity of our experimental setup. In all experiments, the tracking procedure for the last linear layer was conducted using full-batch gradient descent, exactly matching Eq. (8). The mention of "SGD" was intended in the broader sense of first-order gradient methods, though we acknowledge it is imprecise and will update Section 3.2 to specify full-batch gradient descent.
Additionally, RPS-TrReg and RPS-l2 use backtracking line search, which applies a dynamic step size for each update with λ=0.01. TracIn and NTK-Trac use a learning rate of 5×10−3, and RPS-LJE uses 1×10−5 following the original paper. All tracking procedures are conducted for 30 steps. We omitted error bars from the figure to avoid visual clutter given the density of the plotted curves. However, as shown in the table in our response to Weakness 2.1, our results are stable with low variance. We will include a detailed table with full standard deviations in the Appendix of the revised version.