Why Can Accurate Models Be Learned from Inaccurate Annotations?¶
Conference: ECCV2026
Paper: ECCV 2026 Official Page
Area: Optimization & Theory (learning theory under noisy annotation)
Keywords: label noise, partial label learning, singular value decomposition, principal subspace, training-free post-processing
TL;DR¶
The paper turns "why can accurate models be learned from inaccurate annotations?" into an analyzable question: label inaccuracy only deposits noise in the lower singular components of the classification weight matrix, and at moderate inaccuracy rates the principal subspace barely rotates, so task-relevant knowledge survives. This is argued empirically through singular-value spectra and Grassmann subspace similarity and theoretically through a Davis-Kahan angle bound, and it motivates LIP — a training-free post-processing plug-in that keeps the top-\(k\) singular components and re-estimates the flattened tail singular values from the training set — which improves 11 baselines on 6 real-world partial-label datasets and on CIFAR/CUB noisy settings.
Background & Motivation¶
Inaccurate annotations are ubiquitous in practice: image labels come from crowdsourcing, names in video face naming must be scraped from subtitles, and medical or fine-grained labels need experts, so perfectly labeled data is prohibitively expensive. Deep learning therefore routinely operates under two forms of inaccurate annotation — noisy labels, where the true label is dropped or replaced, and partial labels, where an image is paired with a candidate set containing exactly one true label. A phenomenon keeps being observed but has never been explained head-on: models trained on such data still predict correctly, and in some settings approach models trained with ground-truth labels. The question of why has been sidestepped — the past decade of work answers "how to train," not "why learning still works."
Prior methods fall into two lines, neither of which touches that question. Noise-robust training designs losses that are insensitive to corruption (symmetric cross-entropy), dynamically reweights samples by model confidence, relies on early stopping to halt before noise memorization, or explicitly estimates a noise transition matrix to correct labels. Partial-label learning instead treats the true label as a latent variable and disambiguates it via EM, graph-based label propagation, contrastive prototypes, or a complementary classifier (PRODEN, PICO, PL-CL, PL-AGGD). These mechanisms differ, but they share an untested premise: that the correct information is still somewhere in the model and only needs to be preserved. Testing that premise requires a comparable carrier of task knowledge, and the most concrete one is the classification weight matrix — the optimum of a stand-alone method, or the final classification FC layer of a deep network. The question then becomes far more specific: writing the clean-label weights as \(W\) and the inaccurate-label weights as \(W'\), where exactly do they differ, and under what conditions does that difference stay controllable?
The paper answers in two steps — diagnose first, then explain, then reverse the explanation into a method. The diagnosis localizes the damage: the perturbation induced by label inaccuracy accumulates in the lower singular components, whose spectrum is flattened toward zero, while the head of the spectrum (the principal subspace) is essentially untouched at moderate inaccuracy. The theory models inaccurate annotation as an additive perturbation of the clean label matrix, \(Y=G+M\), derives a bound on the perturbation norm that grows only as \(\sqrt{p}\), and propagates it to the principal-subspace rotation angle via the Davis-Kahan sine theorem. Core idea: a model can extract correct information from inaccurate annotations because, within a certain range, label inaccuracy does not significantly move the principal subspace learned from clean data — the task-relevant information stays intact in the top singular components of the weight matrix, so it can be kept by truncation and the noise-flattened tail singular values can be re-estimated from the training set.
Method¶
Overall Architecture¶
The paper reads as "diagnose → explain → prescribe." In the diagnostic stage, PRODEN with a ResNet-34 backbone is trained on clean CIFAR-100 to obtain the clean weights \(W\); labels are then flipped with probability \(p\) (each sample keeps at least one label) to train \(W'\), and both matrices are factorized by SVD and compared from two angles — how the singular-value spectrum itself changes with \(p\), and how similar the subspaces spanned by the top-\(i\) columns of one weight and the top-\(j\) columns of the other remain. In the explanatory stage, the closed-form solution of a least-squares classifier serves as the analytically tractable surrogate: inaccurate labels are written as \(Y=G+M\), giving \(W'=W+\Delta W\) and a bound on \(\|\Delta W\|\) in terms of \(p\), after which the Davis-Kahan sine theorem translates the weight-level perturbation bound into a bound on the principal-subspace rotation angle \(\theta\). The prescription is LIP: given any trained classification weight \(W'\) (clean-trained ones included), it factorizes by SVD, keeps the top-\(k\) singular components via PSP to get \(W_k\), re-solves the remaining singular values \(\Sigma_l^*\) via LAP while freezing the left and right singular vectors, and reassembles \(W^*=W_k+U_l\Sigma_l^*V_l^{\mathsf{T}}\) as the final classification weight. No training is involved: the input is a training set plus a trained classification head, and the output is a matrix that can directly replace the original weights.
Key Designs¶
1. Reducing label inaccuracy to weight perturbation: a bridge built from the least-squares closed form
To ask whether knowledge survives, one needs a comparable object with a closed form, and a whole deep network is analytically intractable. The paper restricts attention to the final classification head and abstracts its objective as a regularized least-squares problem: given a feature matrix \(\mathbf{X}\in\mathbb{R}^{n\times q}\) and a clean label matrix \(\mathbf{G}\in\{0,1\}^{n\times l}\), minimize \(\|\mathbf{X}\mathbf{W}-\mathbf{G}\|_F^2+\lambda\|\mathbf{W}\|_F^2\), whose closed-form solution is \(\mathbf{W}=\mathbf{K}^{-1}\mathbf{X}^{\mathsf{T}}\mathbf{G}\) with \(\mathbf{K}=\mathbf{X}^{\mathsf{T}}\mathbf{X}+\lambda\mathbf{I}\). The abstraction is justified because an FC classification layer selects weights precisely to minimize \(\|\mathbf{W}^{\mathsf{T}}\mathbf{x}-\mathbf{y}\|^2\), where \(\mathbf{x}\) are backbone features and \(\mathbf{y}\) one-hot labels. Inaccurate annotation is then modeled as a perturbation of the label matrix: \(\mathbf{Y}=\mathbf{G}+\mathbf{M}\) with \(\mathbf{M}\in\{-1,0,1\}^{n\times l}\) and the constraint \(\mathbf{M}+\mathbf{G}\ge 0\), so that a \(-1\) can only occur where the entry was originally \(1\) — a removed label must have been an assigned one, and every sample keeps at least one label. Substituting into the closed form gives
which unifies both forms of inaccuracy: \(\mathbf{M}_{ij}=+1\) means an extra wrong label (the partial-label case) and \(\mathbf{M}_{ij}=-1\) means a dropped true label (the noisy-label case). The effect of annotation errors on the weights thus becomes a purely additive perturbation \(\Delta\mathbf{W}\), and every subsequent analysis revolves around it.
2. Spectral diagnosis: locating where the damage happens via the singular-value spectrum and subspace similarity
With the pair \(W\) and \(W'\) in hand, the paper examines the difference from two complementary angles. The first is the singular-value spectrum: as \(p\) grows from 0 to 0.4, the leading singular values coincide with those of the clean weights for \(p\le 0.1\), then collapse rapidly beyond a threshold — the spectrum flattens, the effective rank drops, and the smallest singular values are driven toward zero, since noise preferentially corrupts low-rank directions. The spectrum alone, however, only shows that capacity shrank; it cannot say whether task knowledge survived. The second angle is therefore subspace similarity. Following the normalized Grassmann-distance similarity used in LoRA-style analysis, the overlap between the subspace spanned by the top-\(i\) singular vectors of \(V\) and that spanned by the top-\(j\) of \(V'\) is measured as
where 1 means the two subspaces coincide and 0 means they are unrelated. Computing this for all \(i,j\in[1,60]\) shows that the divergence between the two weights grows with inaccuracy, and that the largest changes occur precisely in the principal subspace (the one corresponding to the largest singular values) — evidence that critical task information is being lost, consistent with the accuracy drop over \(p\). Read in reverse, the same figure carries the key observation: for \(p\le 0.1\) the principal subspace is essentially unaffected, sometimes nearly identical, and the corresponding classification accuracy matches that of the clean weights. This is the diagnosis — the damage is layered, noise deposits in the lower components while the head stays fixed within a range.
3. Theoretical account: a perturbation bound plus Davis-Kahan explains why the principal subspace cannot rotate far
The diagnosis is a phenomenon; the paper then explains why the angle must be small in that range. Norm inequalities decompose the perturbation bound as \(\|\Delta\mathbf{W}\|_F\le\|\mathbf{K}^{-1}\|_2\|\mathbf{X}^{\mathsf{T}}\|_2\|\mathbf{M}\|_F\). Since the feature dimension \(q\) is typically far smaller than \(n\), \(\mathbf{X}\) is full rank, so \(\|\mathbf{K}^{-1}\|_2=1/\lambda_{\min}(\mathbf{X}^{\mathsf{T}}\mathbf{X}+\lambda\mathbf{I})=1/(\lambda_{\min}(\mathbf{X}^{\mathsf{T}}\mathbf{X})+\lambda)\) and \(\|\mathbf{X}^{\mathsf{T}}\|_2=\sigma_{\max}(\mathbf{X})\); and since \(\mathbb{P}(\mathbf{M}_{ij}\ne 0)=p\), the mask has \(pnl\) non-zero entries and \(\|\mathbf{M}\|_F=\sqrt{pnl}\), yielding
The bound is then lifted from the weights to the subspace: by the Davis-Kahan sine theorem the angle \(\theta\) between the principal subspaces satisfies \(\sin\theta\le\|\Delta\mathbf{W}\|_2/\delta\), where \(\delta\) is the gap between the principal singular values of the clean weight (treated as a hyper-parameter); combined with \(\|\Delta\mathbf{W}\|_2\le\|\Delta\mathbf{W}\|_F\),
This yields two layers of conclusion. The first answers the opening question: as long as \(p\) stays in a reasonable range, \(\theta\) is bounded to a small value, the principal subspace does not rotate appreciably, and the information needed for classification is preserved. The second is a readable corollary: \(\theta\) grows only as \(\sqrt{p}\), while the prefactor is governed by data geometry (\(\sigma_{\max}\), \(\lambda_{\min}\), \(n\), \(l\), and the spectral gap \(\delta\)), so the "safe range" is dataset-dependent rather than a universal noise-rate constant. The same inequality incidentally explains two empirical facts: regularization helps because a larger \(\lambda\) directly lowers the upper bound of \(\sin\theta\), and feature normalization pushes \(\sigma_{\max}\) down in the same direction. The appendix also supplies an empirical criterion — a deviation within \(\theta\le30^\circ\) is considered tolerable distortion — and reports that for \(p>0.4\) the principal subspace deviates severely.
4. PSP: keep the principal subspace first, which is where most of the gain comes from
The diagnosis immediately suggests a cheap post-processing move: since the critical information sits in the head, keep the head intact. Factorizing \(W'\) by SVD and reassembling the top-\(k\) components as \(\mathbf{W}_k=\mathbf{U}_k\boldsymbol{\Sigma}_k\mathbf{V}_k^{\mathsf{T}}\) truncates the noise-contaminated tail directions, which amounts to a spectral denoising of the classification weights. \(k\) is LIP's only hyper-parameter; the reported empirical setting is \(k=\lceil 0.8l\rceil\) (\(l\) being the number of classes), and the extremely fast execution of LIP also permits a grid search on the validation set. Ablations show this step carries the bulk of the gain: attached to PL-CL it lifts the average over the six real-world datasets from 49.4 to 51.0, already most of LIP's total 52.2.
5. LAP: instead of discarding the tail, re-estimate the singular values that inaccuracy flattened
Truncation is safe but throws the tail away wholesale, whereas work such as SVDiff has shown that bottom singular components also carry information for the weight matrix itself; the near-zero singular values in Fig. 1 look less like intrinsic noise than like components suppressed by label noise. LAP therefore keeps those directions and re-estimates only their strength: with \(\mathbf{U}_l,\mathbf{V}_l\) (the left and right singular vectors not retained) frozen, it returns to the training data and re-solves the singular values,
Setting the gradient with respect to each \(\sigma_j\) to zero gives an element-wise closed form,
(the matrix form ⚠️ refer to the original paper — that equation is corrupted in the cached text). Because the directions are frozen, no new subspace can be introduced; only the weight each tail direction receives changes, which makes better use of the matrix's effective rank. The ablation shows LAP's contribution is small but consistent (PL-CL average 51.0 → 52.2; PRODEN on CUB-200 at \(p=0.04\), 74.41 → 74.51). PSP is responsible for "not being polluted," LAP for "not wasting capacity."
Loss & Training¶
LIP has no training procedure of its own: PSP is one SVD plus truncation with complexity \(O(\min(q^2l,ql^2))\), LAP is one closed-form least-squares solve with complexity \(O(nql)\), giving \(O(nql+\min(q^2l,ql^2))\) overall. Measured runtimes are 9.98 ms on CIFAR-100 and 16.13 ms on CUB-200, negligible against hours of training, so it can serve as a one-shot post-processing step on any classifier. Experimental settings: PICO and PRODEN both use ResNet-34; PRODEN runs 500 iterations on both CIFAR-100 and CUB-200, PICO runs 500 iterations on CUB-200 and 800 on CIFAR-100; CIFAR-100N uses ResNet-50 with SGD (momentum 0.9) for 300 epochs; real-world datasets use ten runs of 50%/50% random splits reporting mean and standard deviation, while CIFAR-100 and CUB-200 use five runs.
Key Experimental Results¶
Main Results¶
The evaluation covers three forms of inaccurate annotation: six real-world partial-label datasets (FG-NET, Lost, MSRCv2, Mirflickr, Soccer Player, Yahoo!News, where the *-marked variants additionally remove the true label of 10% of samples), synthetic noisy labels (symmetric and asymmetric noise on CIFAR-100), and synthetic partial labels (false positive labels generated on CIFAR-100 and CUB-200 by flipping negative labels with probability \(p\)).
Table 1 summarizes LIP attached to seven partial-label baselines (entries are "baseline → +LIP", with the paper's average relative improvement in the last column):
| Method | Lost* | MSRCv2* | Soccer Player* | Yahoo!News* | Avg. ↑Ratio |
|---|---|---|---|---|---|
| PL-CL → +LIP | 68.3 → 72.8 | 45.1 → 52.1 | 52.1 → 53.6 | 60.9 → 63.9 | 5.7% |
| PL-CGR → +LIP | 71.3 → 74.3 | 49.7 → 51.6 | 53.3 → 55.9 | 61.6 → 65.1 | 4.6% |
| DPCLS → +LIP | 69.6 → 72.7 | 53.6 → 54.5 | 54.7 → 55.7 | 61.9 → 63.8 | 3.3% |
| PL-AGGD → +LIP | 68.2 → 72.9 | 45.0 → 52.3 | 51.7 → 53.6 | 60.7 → 63.9 | 6.3% |
| SURE → +LIP | 68.3 → 70.6 | 43.8 → 48.1 | 51.2 → 51.8 | 59.1 → 61.6 | 3.7% |
| LALO → +LIP | 66.7 → 70.9 | 44.1 → 51.3 | 51.5 → 53.5 | 59.3 → 63.4 | 7.0% |
| PL-LEAF → +LIP | 65.1 → 69.4 | 45.7 → 48.3 | 51.1 → 51.8 | 60.0 → 61.7 | 3.9% |
Table 2 covers the noisy-label setting on CIFAR-100 with two baselines (PLS and AGCE), where the absolute gain under high noise rates is actually larger:
| Method | Asym. 10% | Asym. 20% | Asym. 40% | Asym. 50% | Sym. 10% | Sym. 20% | Sym. 40% | Sym. 80% |
|---|---|---|---|---|---|---|---|---|
| PLS → +LIP | 79.54 → 79.75 | 73.90 → 74.34 | 53.17 → 54.42 | 34.51 → 38.49 | 79.65 → 79.87 | 79.13 → 79.69 | 77.27 → 77.38 | 45.92 → 47.43 |
| AGCE → +LIP | 67.32 → 67.56 | 63.57 → 63.99 | 47.25 → 48.79 | 27.90 → 28.25 | 67.61 → 67.98 | 64.75 → 64.93 | 59.78 → 59.91 | 24.05 → 25.43 |
Table 3 covers partial labels (CIFAR-100 with \(p\) = 0.05/0.1/0.2 and CUB-200 with 0.02/0.04/0.06) and also reports the clean-training references:
| Method | CIFAR-100 p=0.05 | p=0.1 | p=0.2 | CUB-200 p=0.02 | p=0.04 | p=0.06 |
|---|---|---|---|---|---|---|
| PICO → +LIP | 73.51 → 73.83 | 72.18 → 72.43 | 71.62 → 71.98 | 72.56 → 72.90 | 72.27 → 72.56 | 71.88 → 72.26 |
| PRODEN → +LIP | 77.12 → 77.25 | 76.04 → 76.42 | 56.83 → 56.90 | 74.53 → 74.73 | 74.36 → 74.51 | 72.01 → 72.33 |
| Clean reference | 73.88 (PICO) / 77.31 (fully supervised), 74.09 / 77.48 with LIP | — | — | 76.00 (PICO) / 75.70 (fully supervised), 76.56 / 75.91 with LIP | — | — |
The PRODEN-CIFAR rows of Table 3 reproduce the accuracies of the diagnostic experiment (77.12 / 76.04 / 56.83), confirming that the spectral observation and the classification behavior come from the same runs.
Ablation Study¶
| Config | PL-CL avg. over 6 datasets | PRODEN CIFAR-100 (p=0.05) | PRODEN CIFAR-100 (p=0.2) | PRODEN CUB-200 (p=0.04) |
|---|---|---|---|---|
| Baseline (no PSP, no LAP) | 49.4 | 77.12 | 56.83 | 74.36 |
| PSP only | 51.0 | 77.21 | 56.89 | 74.41 |
| PSP + LAP (full LIP) | 52.2 | 77.25 | 56.90 | 74.51 |
Key Findings¶
- PSP is the main contributor and LAP a consistent small addition. Removing LAP already retains most of the gain (PL-CL 49.4 → 51.0, i.e. 1.6 of the total 2.8), with LAP supplying the rest; the two modules complement each other on every dataset, matching their "keep the head, fix the tail" division of labor.
- Heavier noise produces the pattern the theory predicts. On weak baselines with high noise rates, LIP's absolute gain is largest (PLS asymmetric 50%: 34.51 → 38.49, +3.98; symmetric 80%: 45.92 → 47.43; AGCE symmetric 80%: 24.05 → 25.43), indicating that the flattened tail still holds usable information. On PRODEN at \(p=0.2\), however, the gain shrinks to 56.83 → 56.90 (+0.07): at that level the principal subspace itself is damaged, and post-processing has nothing left to repair — exactly the theoretical expectation.
- LIP also helps on clean data, which goes beyond the noise-robustness intuition: fully supervised CIFAR-100 improves 77.31 → 77.48, PICO's clean setting 73.88 → 74.09, and the CUB-200 clean references 76.00 → 76.56 and 75.70 → 75.91. Under normal training the tail singular values are likewise under-utilized (the two clean references use different protocols, so their absolute values should not be compared directly).
- Consistent gains on real-world noisy datasets: with PLS, CIFAR-10N improves 85.99 → 86.69, CIFAR-100N 55.54 → 56.46, and Clothing1M 47.59 → 48.36.
- Robust to \(k\): on CUB-200 with PRODEN, performance rises with \(k\) and then stabilizes; beyond \(k>100\) LIP is consistently better than PRODEN alone. With \(k=\lceil0.8l\rceil\), the 200-class CUB-200 lands at \(k=160\), inside the stable region — the heuristic and the sensitivity curve agree.
- Backbone-agnostic: replacing PRODEN's backbone with ViT-B reproduces the same trend of subspace similarity versus \(p\) as ResNet-34, and LIP still improves results (81.14 → 81.89, 80.39 → 81.07, 62.24 → 63.13), supporting the claim that the conclusion stems from classification-head geometry rather than convolutional structure.
- Statistically significant: every baseline pairing in Table 1 passes a pairwise t-test at the 0.05 significance level.
Highlights & Insights¶
- Turning "why" into a computable question. Instead of discussing generalization directly, the paper moves annotation errors into a weight perturbation, so "is the knowledge still there" becomes an angle between singular subspaces that Davis-Kahan can bound. The whole route transfers to other "training looks broken but results are fine" phenomena, such as distribution drift during PEFT fine-tuning or weight-drift diagnosis in continual learning.
- Diagnosis and method correspond strictly. The diagnosis says "damage in the tail, head intact," so the method keeps the head and fixes the tail; LAP re-estimates only singular values rather than directions precisely because frozen directions cannot introduce a new subspace. The method is a direct translation of the conclusion, not an ad-hoc design.
- A counter-intuitive by-product. LIP also helps on clean data, implying that "under-utilized tail singular values" is not a pathology specific to noisy annotation but a general property of least-squares-style classification heads. This explains why a plug-in designed purely for noise robustness still works under full supervision.
- A new reading of regularization. Since \(\lambda\) and \(\sigma_{\max}\) sit in the denominator of the \(\sin\theta\) bound, regularization (larger \(\lambda\)) and feature normalization (lower \(\sigma_{\max}\)) can be read as lowering the upper bound on how far the principal subspace can be rotated, rather than as a vague "prevention of overfitting."
- A reusable trick. LIP's "SVD → freeze directions → closed-form re-estimation of singular values from the training data" is a general weight post-processing template that is nearly free (milliseconds, no gradients) and applies to any model ending in a linear classification head, including already-deployed models that cannot be retrained.
Limitations & Future Work¶
- The analysis is confined to the final classification head. Intermediate layers receive only an abstract argument (approximating each layer as a linear input-to-output mapping) with no verifiable layer-wise experiment; the authors suggest training clean and noisy networks and comparing the SVD structure of corresponding intermediate weights, but the main text does not do this. A real gap remains: if noise mainly corrupts backbone features rather than the head, this conclusion does not cover it.
- The "safe range" cannot be computed directly. The gap \(\delta\) is treated as a hyper-parameter and \(\sigma_{\max}\), \(\lambda_{\min}\) depend on the dataset, so no a priori noise-rate threshold follows from the theory; both \(\theta\le30^\circ\) and the \(p>0.4\) severity claim are empirical. For a new dataset, whether LIP is worth applying still has to be decided experimentally.
- LAP's refit still uses the noisy labels \(\mathbf{Y}\). Because the directions are frozen it cannot repair a principal subspace that has already drifted, which directly explains the mere +0.07 for PRODEN at \(p=0.2\) — the part needing repair there is the head, while LIP only touches the tail. Replacing \(\mathbf{Y}\) with a debiased or reweighted label estimate (down-weighting suspected errors) is a natural improvement direction.
- The choice of \(k\) lacks theoretical guidance. The heuristic \(k=\lceil0.8l\rceil\) is tied to the number of classes, whereas in theory \(k\) should relate to the singular gap \(\delta\); estimating \(\delta\) from data and selecting \(k\) automatically is an unsolved interface problem.
- The constructed noise is fairly idealized. The main experiments use flip-based synthetic noise, and real-world datasets are reported only through mean accuracy with uncontrollable noise rates and types. Real annotation errors are typically instance-dependent and structured, whereas the \(\mathbf{M}\) model assumes each entry errs independently with probability \(p\); whether that assumption holds in practice is not tested.
- LIP acts on a single trained weight matrix and offers nothing for training dynamics (when memorization happens, when to stop early), so it can only complement, not replace, methods that intervene during training.
Related Work & Insights¶
- vs noisy label learning (robust losses such as SCE, sample reweighting, early stopping, transition matrices, DivideMix / Co-teaching): they fight noise during training by changing the loss, the sampling weights, the label estimates, or the stopping point; this paper never intervenes in training and instead performs one spectral post-processing step on the learned classification weights. The difference is explanatory: the shared premise of "the correct information is still there" is precisely what this paper verifies and bounds. In practice the two are orthogonal and stack.
- vs partial-label disambiguation methods (PRODEN, PICO, PL-CL, PL-AGGD, PL-CGR): they treat the true label as a latent variable and lock it down during training, which is a "how to train" answer; LIP assumes training has finished and edits only the final weights, a "what to do after training" answer. Applying LIP on top of seven such baselines and improving all of them is itself the most direct evidence of orthogonality — it adds a layer of gain without redesigning the disambiguation training.
- vs weight-spectrum analyses such as LoRA, SVDiff, and FLoRA: the subspace-similarity metric follows LoRA's normalized Grassmann distance, and LAP's emphasis that bottom singular components matter echoes SVDiff's observation. The goals differ: those works target parameter-efficient fine-tuning or compression (how to stay accurate under a low-rank constraint), while this paper targets robustness attribution (which components label noise pushes power into) — the same tool applied to a different question.
- vs generalization-bound theory: this paper derives a subspace-angle bound, not a generalization-error bound, so it explains why task information remains in the weights rather than how large the generalization error is. ⚠️ This is my own assessment: linking the \(\sin\theta\) bound to downstream accuracy currently rests on experimental correlation (accuracy in Fig. 1 and angle in Fig. 2 move together), not on a strict derivation chain.
Rating¶
- Novelty: ⭐⭐⭐⭐ The first work to directly ask why inaccurate annotations still yield correct knowledge, with empirical and theoretical evidence on both sides
- Experimental Thoroughness: ⭐⭐⭐⭐ 11 baselines × 6 real-world datasets plus synthetic noise/partial labels, ablations, sensitivity, and architecture validation, though the analysis itself centers on a single backbone and synthetic noise
- Writing Quality: ⭐⭐⭐⭐ The diagnosis-theory-method loop is clear and every claim maps to an experiment, though key assumptions (FC-layer abstraction, \(\delta\) as a hyper-parameter) are mainly deferred to the appendix
- Value: ⭐⭐⭐⭐ Training-free, millisecond-scale, and pluggable, with a reusable framework explaining why regularization helps