When a classifier is trained on long-tailed data, where a few classes have thousands of examples and many have only a handful, two kinds of training examples end up with high loss. The first are genuine examples of rare classes, which the model gets wrong because it has learned to favor frequent classes. The second are mislabeled examples, which the model gets wrong because their labels are false. Tilted empirical risk minimization (TERM) reweights each training example by the exponential of its loss. With a positive tilt it emphasizes high-loss examples, and with a negative tilt it suppresses them. Neither setting works when both kinds of high-loss example are present. A positive tilt amplifies the mislabeled examples. A negative tilt throws away the rare classes. Tail classes are where both problems concentrate: they have the fewest examples, so a few corrupted labels make up a large share of each class.
This note proposes a simple change. Before tilting, subtract the part of each example's loss that comes only from class frequency. Concretely, compute the tilt weights on the logit-adjusted loss of Long-tail learning via logit adjustment, which adds each class's log prior to the logits during training. After this correction, a clean tail example is no longer high-loss just because its class is rare. The losses that stay high within a class should then be mostly label noise. A negative tilt inside each class can then act as a noise filter without consuming the tail. An optional positive tilt across classes then targets the worst-performing classes.
Core Claim
The claim is that the order of operations matters. Hierarchical tilting applied to raw cross-entropy confuses rarity with corruption. The same tilting applied to prior-corrected losses separates them. Let $\pi_y$ be the training frequency of class $y$, $f(x)$ the model's logits, and $\ell_i$ the cross-entropy of $f(x_i) + \tau \log \pi$ against label $y_i$, with $\tau$ the logit-adjustment strength. The proposed objective is
$$R(\theta) = \frac{1}{t_{\text{out}}} \log \frac{1}{K}\sum_{y=1}^{K} \exp\big(t_{\text{out}} R_y(\theta)\big), \qquad R_y(\theta) = \frac{1}{t_{\text{in}}} \log \frac{1}{n_y}\sum_{i: y_i = y} \exp\big(t_{\text{in}}\, \ell_i(\theta)\big).$$
Here $K$ is the number of classes, $n_y$ is the number of examples labeled $y$, $t_{\text{in}} < 0$ is the within-class tilt that suppresses outliers, and $t_{\text{out}} \ge 0$ is the across-class tilt.
The objective reduces to known methods at its limits. At $t_{\text{in}} \to 0$ and $t_{\text{out}} \to 0$ it becomes a class-balanced logit-adjusted loss. The class-balanced average and the logit offset both target balanced error, so this limit is a balanced-error method that is not exactly Menon et al.'s frequency-weighted logit-adjusted loss. At $\tau = 0$ it becomes the Hierarchical TERM of Tilted Empirical Risk Minimization. The new element is the $\tau \log \pi$ term inside the inner tilt. A secondary element is a deferred schedule: run plain logit-adjusted training first and switch on the tilt only once losses carry information about label quality. This follows the deferral argument of Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss.
Relation to Prior Work
Three papers supply the method.
- TERM and Hierarchical TERM. Tilted Empirical Risk Minimization introduced the tilted objective and its hierarchical form. It showed on the binary HIV-1 dataset, which has class imbalance and 30% label noise, that an outer positive tilt with an inner negative tilt reaches 90.1% accuracy. Its imbalance experiments stay at the scale of MNIST and HIV-1, and its losses are never corrected for class priors. We reuse its objective and change what it is applied to.
- Logit adjustment. Long-tail learning via logit adjustment supplies the prior correction. It proves that adding $\tau \log \pi_y$ is consistent for balanced error. Our use of it is different: Menon et al. use the correction to fix the decision rule, and we use it to make per-example losses a trustworthy signal of label quality.
- Deferred re-weighting. The LDAM paper, Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss, showed that imbalance corrections applied from the first step hurt feature learning. That result motivates our warm-up period.
Two papers constrain the design. Tilted Cross-Entropy (TCE): Promoting Fairness in Semantic Segmentation is the closest existing use of tilting against imbalance. It tilts class-level losses for segmentation on Cityscapes and reports 78.16% mean intersection-over-union (mIoU), against 75.79% for plain cross-entropy. It applies only a positive class-level tilt from epoch zero, with no within-class noise suppression and no prior correction. If our outer tilt alone explained the gains, our contribution would reduce to TCE moved to classification, so the ablations below isolate the inner, prior-corrected tilt. Influence of minority class instance types on SMOTE imbalanced data oversampling shows that minority examples include safe, borderline, rare, and outlier types. Our filter must therefore be judged on whether it keeps borderline and rare tail examples while dropping the outliers. Improving average accuracy is not enough.
A third group sets the objective we report against. Balancing the Scales identifies a "margin paradox": optimizing overall accuracy calls for larger margins on majority classes, while optimizing balanced error calls for larger margins on minority classes. We make no claim of standard-accuracy optimality. We target balanced error and worst-class error, and we report overall accuracy to make the trade-off visible. Prior2Posterior shows that the prior a trained model has actually absorbed can differ from the empirical frequencies $\pi_y$. That suggests a variant that corrects losses with the estimated absorbed prior instead of $\pi_y$, and we test it.
Research Program
Diagnostic study. The first part is cheap and decides whether the idea is worth pursuing. Train a ResNet on CIFAR-100-LT, the long-tailed version of CIFAR-100, at imbalance factors 100 and 200. Inject two kinds of label noise: symmetric flips at 20% and 40%, and asymmetric flips that move head-class images into tail labels. The asymmetric case matters most because it makes corrupted examples a large share of each tail class. At several checkpoints, measure how well the loss separates noisy from clean examples, using the area under the ROC curve (AUROC) computed separately for head, middle, and tail classes. Compare raw cross-entropy with prior-corrected loss. The hypothesis predicts that raw-loss AUROC on tail classes is much lower than on head classes, and that prior correction closes most of that gap. If prior correction does not raise tail AUROC, the method has no basis. We would report that negative result and stop.
Method study. The second part trains the full objective with the deferred schedule. Mini-batches use the class-stratified stochastic estimator of TCE, so each class's tilted loss is estimated from its own running statistics. The tilts $t_{\text{in}}$ and $t_{\text{out}}$ are tuned by grid search on a held-out set that has balanced classes and clean labels. We also test a fixed rule that sets $t_{\text{in}}$ from each class's estimated noise share. This rule addresses the known hyperparameter-sensitivity problem of tilting.
Analysis. For every trained model we track how much training weight lands on clean tail examples compared with noisy tail examples. This measures directly whether the filter avoids consuming the tail. We also measure how far the classifier vectors depart from a symmetric equiangular arrangement. On the Implicit Geometry of Cross-Entropy Parameterizations for Label-Imbalanced Data predicts minority collapse under plain cross-entropy, and the question is whether prior-corrected tilting avoids it.
Baselines
Plain ERM. Standard cross-entropy is the do-nothing reference. LDAM reports 29.64% error for ERM on CIFAR-10-LT at imbalance 100 and 42.86% top-1 error on iNaturalist 2018.
Logit-adjusted loss. The logit-adjusted loss of Long-tail learning via logit adjustment is the strongest balanced-error competitor and the nearest point inside our own family, at zero tilt. It reports 33.64% balanced error on iNaturalist 2018, and 31.56% when combined with adaptive margins. We will rerun it, with and without class-balanced averaging, using our backbone, epoch count, and augmentation, and give it the same tuning budget for $\tau$ as our method receives for $(t_{\text{in}}, t_{\text{out}})$. On clean data we expect at best parity. Beating it means lower balanced error and worst-class error on the noisy long-tailed settings at matched compute, without losing more than a small margin on clean data.
LDAM-DRW. LDAM-DRW is the main margin-based competitor. It reports 22.97% error on CIFAR-10-LT at imbalance 100 and 32.00% top-1 error on iNaturalist 2018. On CIFAR-100-LT at imbalance 100, CVS+ADRW from A Unified Perspective for Loss-Oriented Imbalanced Learning via Localization reports balanced accuracies of 44.4%, 48.3%, and 54.6% under its three training protocols. We will match one of these protocols. We found no result on Lacuna for any of these methods on CIFAR-LT with injected label noise, so all noisy-setting baseline numbers will come from our own reruns.
Hierarchical TERM on raw losses. Li et al.'s objective with $\tau = 0$ is the ablation that isolates the key ingredient. It is identical to our method except for the prior correction.
Outer tilt only. The setting $t_{\text{in}} = 0$ is effectively TCE moved to classification, and it isolates the inner filter.
Simplified training pipeline. The recipe of Simplifying Neural Network Training Under Class Imbalance uses small batches, joint self-supervision, class-conditional label smoothing, and asymmetric sharpness-aware minimization. It reports a 5.4% to 6.8% error reduction on the 20 smallest CIFAR-100 classes. We run it both as a competitor and with our loss substituted into the recipe.
Why not just…?
Why not just train with logit adjustment and add a standard small-loss filter, keeping the lowest-loss fraction of examples in each class? This is the most serious threat, because it is roughly a hard-threshold version of our inner tilt. If it is applied to prior-corrected losses, it already contains our key idea. If it is applied to raw losses, the diagnostic study predicts it will drop clean tail examples. We will run both versions with the keep fraction tuned on the same budget. If the prior-corrected hard filter matches our smooth tilt, we will say so: the contribution then becomes the finding that prior correction must come before any loss-based noise filtering, which is still useful. The smooth tilt's remaining advantage would be that it does not require choosing a keep fraction per class.
Why not just use Hierarchical TERM as published? It already handles imbalance and noise together on HIV-1. But its inner tilt sees raw losses, which on tail classes are high whether an example is clean or corrupted. The $\tau = 0$ ablation and the weight-on-clean-tail measurement test this directly. We expect the raw version to put noticeably less weight on clean tail examples.
Why not just use LDAM-DRW or the simplified pipeline and ignore noise? On clean benchmarks this may well be enough, and we expect no large gain there. Neither method has a mechanism that separates corrupted tail labels from clean ones. LDAM in particular enlarges margins for tail examples, which includes the corrupted ones. The asymmetric-noise experiments show whether that gap appears in practice.
Why not just apply post-hoc logit adjustment or Prior2Posterior after robust training? Post-hoc corrections change the decision rule but cannot undo the damage when the training signal itself was polluted by amplified noise or by discarded tail examples. We include "negative-tilt TERM followed by post-hoc adjustment" as a baseline to test this.
Decisive Experiments
Testing proceeds in three stages.
Diagnostic gate. The first check is the tail-AUROC comparison described in the Research Program. If prior correction does not clearly raise the separability of noisy from clean labels in tail classes, we stop.
Main comparison. The second stage runs on CIFAR-10-LT and CIFAR-100-LT at imbalance factors 100 and 200, under clean labels, symmetric noise at 20% and 40%, and head-to-tail flips. Metrics are balanced error, worst-class error, mean error on the 20 smallest classes, and overall accuracy, each averaged over three seeds. The key comparison is against the $\tau = 0$ ablation and the prior-corrected hard filter. The method succeeds if it beats logit adjustment and raw Hierarchical TERM on noisy tail error, and is within noise of them on clean data.
Real-data check. The third stage evaluates on iNaturalist 2018 and ImageNet-LT, where any label noise is natural rather than injected. Here the target is to match or improve on the 33.64% balanced error that Menon et al. report for logit adjustment on iNaturalist 2018. We also rerun HIV-1 so that results connect directly to the original TERM paper.
References
- Long-tail learning via logit adjustment
- Tilted Empirical Risk Minimization
- Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss
- Tilted Cross-Entropy (TCE): Promoting Fairness in Semantic Segmentation
- Influence of minority class instance types on SMOTE imbalanced data oversampling
- Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data
- Prior2Posterior: Model Prior Correction for Long-Tailed Learning
- On the Implicit Geometry of Cross-Entropy Parameterizations for Label-Imbalanced Data
- A Unified Perspective for Loss-Oriented Imbalanced Learning via Localization
- Simplifying Neural Network Training Under Class Imbalance