Curriculum learning trains a model on easy examples first and admits harder ones as training goes on. Its most common automatic form, self-paced learning, lets the model decide what counts as "easy" by admitting any example whose current loss falls below a rising threshold. On class-imbalanced data this is a trap. A network trained on data dominated by a few frequent classes becomes overconfident on those classes, so their examples have low loss and pass the threshold, while rare-class examples stay out. The diversity-aware variants built to stop curricula from fixating on redundant easy examples make things worse. They reward picking examples from many different clusters, and frequent classes supply most of the clusters. The curriculum ends up amplifying the majority-class bias it was never designed to handle.
We propose to split the curriculum's two decisions and correct each one separately. The first decision is how many examples of each class to admit. We make it by thresholding a loss from which the model's own learned class prior has been subtracted, so that head-class examples no longer look easy just because the model favors head classes. The second decision is which examples of a class to admit. We make it with a diversity bonus computed only across sub-clusters inside that class, normalized so every class gets the same total bonus. The result is a self-paced curriculum for deep long-tailed recognition whose pacing cannot be captured by the majority. We will test whether it yields better features than simply rebalancing the loss or correcting the outputs after training.
Core Claim
Our first claim is diagnostic. Standard self-paced learning, and its diversity-regularized form, admits frequent-class examples disproportionately on long-tailed benchmarks, and this admission skew predicts tail-class accuracy. Our second claim is constructive. If admission is driven by prior-corrected loss and diversity is confined within classes, the curriculum admits a class-balanced but internally varied set. The payoff we care about is better representations for tail classes, and it should survive the strongest post-hoc correction available. This last condition is what separates a genuine training effect from a disguised shift of the output prior.
Relation to Prior Work
The diversity mechanism comes from Self-Paced Learning with Diversity (SPLD), which adds a negative $\ell_{2,1}$ penalty over data groups. Under this penalty, the first examples taken from a group receive a bonus that shrinks as more are taken from the same group. SPLD beat plain self-paced learning on TRECVID MED13, reaching 12.1 mean average precision against 9.6. Its groups are treated symmetrically, however. On imbalanced data, a rare class occupying one cluster has its second and third examples penalized as "redundant," while a frequent class spread over dozens of clusters collects dozens of first-pick bonuses.
Two related methods share this blind spot. Balanced Self-Paced Learning for Generative Adversarial Clustering Network uses an exclusive-lasso penalty to balance selection across latent clusters. Self-Paced Multi-Label Learning with Diversity spreads selection across labels. Both balance over clusters or labels, not over classes weighted by how many sub-clusters each class contains.
Minimax Curriculum Learning instead selects hard but diverse subsets with submodular functions. It reports that plain self-paced learning is worse than random minibatches on CIFAR-10 (21.14% versus 18.52% test error) and that its own method reaches 12.87%. That result motivates us, because it shows loss-only pacing can actively hurt. It also constrains us, because any curriculum we propose must beat random minibatches, not just plain self-paced learning.
The correction to admission comes from Prior2Posterior (P2P). P2P shows that the class prior a trained network actually internalizes, which it estimates by averaging predicted posteriors, differs from raw class frequencies. Correcting outputs with this estimated prior lifts CIFAR100-LT accuracy at imbalance factor 200 from 43.39% to 52.92%. P2P applies the correction once, after training. We apply a running estimate of the same quantity during training, inside the admission rule, which P2P never does.
The within-class grouping reuses the idea of Modeling Skewed Class Distributions by Reshaping the Concept Space, which splits classes into sub-clusters and trains on the sub-labels. That work does this as a static preprocessing step. We instead recompute sub-clusters from the network's current features and use them only to decide which examples to admit, never as training labels.
Several existing methods already sit at the intersection, and we differ from each. Dynamic Curriculum Learning moves from natural to class-balanced sampling on a fixed epoch schedule, with no feedback from the model and no structure inside classes. CUDA paces augmentation strength per class and changes no sample weights. Self-paced Ensemble and DuBE stratify majority examples by a scalar hardness score inside tree ensembles for mostly binary tabular data. CASED is a hand-designed schedule that starts from minority-only patches for binary detection. None of these corrects a self-paced admission threshold for the learned prior, and none decouples per-class quantity from within-class diversity.
Method
Let $\hat\pi$ be the running estimate of the model's effective class prior: the exponential moving average, over recent minibatches, of the softmax outputs averaged across training examples. Let $z_i$ be the logits for example $i$. The admission loss is the cross-entropy computed on shifted logits:
$$\tilde L_i = L\big(y_i,\ z_i - \tau \log \hat\pi\big),$$
where $\tau$ is a correction strength, with $\tau=1$ as the default.
Within each class $c$, we run k-means on the current penultimate-layer features every $E$ epochs to form $K_c$ sub-clusters $G_{c,1},\dots,G_{c,K_c}$. The inclusion weights $v$ then solve:
$$\min_{v\in[0,1]^n}\ \sum_i v_i \tilde L_i - \lambda\sum_i v_i - \gamma \sum_c \frac{1}{\sqrt{K_c}}\sum_{k=1}^{K_c}\big\|v^{(c,k)}\big\|_2 .$$
Here $\lambda$ is the pacing threshold, which rises over training. The vector $v^{(c,k)}$ holds the weights of the examples in sub-cluster $G_{c,k}$. The last term is SPLD's negative $\ell_{2,1}$ bonus with a per-class factor $1/\sqrt{K_c}$. If each of a class's $K_c$ sub-clusters has a share $m$ of its examples selected, the class's total bonus is $\gamma\sqrt{m K_c}$, which depends only on how many examples the class has admitted, not on how many clusters it was split into. The factor therefore keeps a many-cluster head class from collecting more bonus than a one-cluster tail class at the same admission level.
SPLD's sort-and-threshold solver carries over unchanged. Within each sub-cluster, the $j$-th lowest-loss example is admitted when
$$\tilde L < \lambda + \frac{\gamma}{\sqrt{K_c}}\cdot\frac{1}{\sqrt{j}+\sqrt{j-1}}.$$
The cost is a per-class sort, the periodic clustering, and a moving average that needs no extra forward passes. This is lighter than CUDA's roughly 1.3x overhead. The network is trained on plain cross-entropy weighted by $v$, which isolates the effect of the curriculum. A variant that also uses the adjusted loss for training is reported separately.
We deliberately keep the prior shift additive and confine it to the admission rule. A Unified Perspective for Loss-Oriented Imbalanced Learning via Localization proves that multiplicative logit margins combined with deferred reweighting loosen generalization bounds. A rising-threshold curriculum is a form of deferred reweighting, so we avoid multiplicative margins in the method itself and test the combination with that paper's loss explicitly.
Research Program
The work has three stages.
The first stage is a measurement study. On CIFAR-10-LT and CIFAR-100-LT, we log the per-class admission rate of plain self-paced learning and SPLD throughout training, alongside the gap between $\hat\pi$ and the empirical class frequencies. We then correlate each class's admission deficit with its final accuracy. If admission does not skew toward head classes, the motivating mechanism is false and the project stops or pivots.
The second stage builds and ablates the method. Each ablation tests one ingredient: the prior correction, the within-class restriction on diversity, the $1/\sqrt{K_c}$ normalization, and the use of feature-based clusters rather than random groups.
The third stage asks what the curriculum changes inside the network. We retrain a class-balanced linear classifier on frozen features and apply P2P to every model before comparing. Both steps are needed to separate representation gains from shifts in the output prior.
Baselines
The naive reference point is ResNet-32 trained with ordinary cross-entropy on CIFAR-10/100-LT (imbalance factors 10, 100, 200), and ResNet-50 on ImageNet-LT and iNaturalist18. CUDA reports 38.7% for plain cross-entropy on CIFAR-100-LT, though the file does not give the imbalance factor for this number. P2P reports a 43.39% baseline at imbalance factor 200.
The curriculum baselines are plain self-paced learning, SPLD with class-agnostic k-means groups, and Minimax Curriculum Learning. The last is the strongest diversity-aware curriculum and the only one known to beat random minibatches.
The strongest imbalance competitors are three:
- CVS + ADRW (Localization): reports 44.4%, 48.3%, and 54.6% balanced accuracy on CIFAR-100-LT at imbalance factor 100 under its three training protocols. We will match its protocol exactly rather than compare across protocols.
- P2P (Prior2Posterior): 52.92% on CIFAR100-LT at imbalance factor 200 and 54.67% top-1 on ImageNet-LT with feature tuning.
- CUDA (CUDA): 42.7% with cross-entropy and 52.3% combined with balanced contrastive learning, imbalance factor not given in the file. On iNaturalist18, CUDA reports a 1.2–1.5% gain over its baselines, but no reported absolute result was found on Lacuna for iNaturalist18 for any method.
We also include DCL's fixed natural-to-balanced schedule, and class-balanced sampling with a logit-adjusted loss. The latter has no paper in the file and will be implemented from its standard definition.
For fairness, all methods share the same backbone, epoch count, augmentation pipeline, and a tuning budget of 20 configurations each on a held-out balanced validation split. Where published numbers were obtained under different protocols, we rerun the baselines rather than quote them. Beating the field means two things: exceeding CVS + ADRW under the matched protocol, and exceeding P2P when P2P is also applied on top of our model. The decisive ablations replace $\hat\pi$ with empirical class frequencies, drop the $1/\sqrt{K_c}$ factor, and use symmetric class-agnostic groups.
Why not just…?
Why not just train with a logit-adjusted or class-balanced loss and skip the curriculum? This is the most serious threat, because such a loss already removes much of the head-class advantage in the loss values themselves. Our claim is that admission control adds a second benefit: tail examples are not diluted early by redundant head examples, and within-class diversity keeps features spread out. The linear-probe comparison settles this. If our method's frozen features give no better tail accuracy than those of a logit-adjusted model, the curriculum adds nothing and we will report exactly that.
Why not just train normally and apply P2P afterward? P2P gains nine points at imbalance factor 200 with no retraining, so a post-hoc fix is a strong bar. But a post-hoc shift cannot repair features learned under skewed gradients. The Localization paper also shows that minority classes stay miscalibrated under global corrections. We will apply P2P to every model. If our advantage vanishes after P2P, then our curriculum only moved the prior and did not improve features.
Why not just use DCL's fixed schedule from natural to balanced sampling? DCL reaches 89.05% mean accuracy on CelebA and 83.7% on RAP, so scheduled rebalancing works. However, its schedule ignores both model feedback and structure inside each class. Running it at matched budget on CelebA and CIFAR-LT, and comparing with our variant that removes the diversity term, isolates whether feedback-driven admission matters on its own.
Why not correct the threshold with raw class frequencies instead of the learned prior? P2P's evidence that the two differ is the reason to expect a difference, but the size of the effect under a curriculum is unknown. The frequency-prior ablation measures it directly. If the two perform the same, the method simplifies and the result is still worth reporting.
Decisive Checks
We will measure three things: balanced accuracy, accuracy separately on head, medium, and tail classes, and per-class expected calibration error. For CelebA, we report mean attribute accuracy.
The project succeeds if four conditions hold together:
- The first-stage study shows admission skewed toward head classes under both plain self-paced learning and SPLD.
- Our method removes that skew.
- Our method exceeds the matched CVS + ADRW baseline and still beats P2P-corrected cross-entropy after P2P is applied to both.
- The frozen-feature linear probe shows higher tail accuracy than the logit-adjusted baseline.
A final robustness check injects 20% and 40% label noise into CIFAR-100-LT. DuBE found that standard hard example mining collapses beyond 20% noise. A loss-threshold curriculum should resist noise, and we will verify that the prior correction does not erase this robustness by admitting mislabeled tail examples early.
References
- Self-Paced Learning with Diversity
- Balanced Self-Paced Learning for Generative Adversarial Clustering Network
- Self-Paced Multi-Label Learning with Diversity
- Minimax Curriculum Learning: Machine Teaching with Desirable Difficulties and Scheduled Diversity
- Prior2Posterior: Model Prior Correction for Long-Tailed Learning
- Modeling Skewed Class Distributions by Reshaping the Concept Space
- Dynamic Curriculum Learning for Imbalanced Data Classification
- CUDA: Curriculum of Data Augmentation for Long-tailed Recognition
- Self-paced Ensemble for Highly Imbalanced Massive Data Classification
- Towards Inter-class and Intra-class Imbalance in Class-imbalanced Learning
- CASED: Curriculum Adaptive Sampling for Extreme Data Imbalance
- A Unified Perspective for Loss-Oriented Imbalanced Learning via Localization