Ω(pₜ, f_c) in Long-Tailed Learning

• Ω(pₜ, f_c) is a closed-form, dual-phase reweighting factor that adjusts per-sample loss based on prediction confidence and empirical class frequency.
• It employs a maximum-entropy margin principle to modulate gradients, balancing exploration for uncertain samples with consolidation for confident predictions.
• Empirical results demonstrate that this function improves tail-class accuracy across benchmarks like CIFAR-100-LT and ImageNet-LT.

The Ω(pₜ, f_c) function is a closed-form, dual-phase, sample-wise reweighting factor introduced in the context of Class Confidence Aware Reweighting (CCAR) for long-tailed learning. Designed to complement existing logit-adjustment and margin-based loss modifications, Ω(pₜ, f_c) adapts each training sample’s loss contribution according to both the model’s predicted probability for the ground-truth class ($pₜ$) and the empirical class frequency ($f_c$), enabling simultaneous emphasis on hard samples and tail-class amplification. The function has emerged as an efficient, bounded, and theoretically principled alternative for mitigating class imbalance, delivering consistent accuracy improvements across a variety of long-tailed benchmarks (Jagati et al., 22 Jan 2026).

1. Definition and Formal Specification

Let $pₜ\in(0,1]$ be the predicted probability for the true class $t$, and $f_c=N_c/N$ be the empirical frequency of class $c$. The function incorporates a confidence pivot $\omega\in(0,1]$ (commonly $\omega=0.75$). The core ingredient is a piecewise, "dual-phase" frequency:

$$f'_c(pₜ) = \begin{cases} f_c, & \text{if } pₜ<\omega \quad\text{("exploration" phase)} \ 1-f_c, & \text{if } pₜ\ge\omega \quad\text{("consolidation" phase)} \end{cases}$$

The CCAR sample weight is then given by:

$$\Omega(pₜ, f_c) = (e - f'_c(pₜ))^{\omega - pₜ} = \exp\big[(\omega - pₜ)\cdot\ln(e - f'_c(pₜ))\big]$$

Introducing $\beta_c(pₜ) := \ln(e - f'_c(pₜ))$ (the adaptive capacity), the expression becomes:

$$\Omega(pₜ, f_c) = \exp\big[\beta_c(pₜ)\,(\omega - pₜ)\big]$$

This construction ensures $\Omega(\omega, f_c) = 1$ for all $f_c$, enforcing continuity at the pivot.

2. Theoretical Motivation and Behavior

Maximum-Entropy Margin Principle

The function arises from a maximum-entropy variational principle. By defining a margin utility $m(pₜ) = \omega - pₜ$, the problem

$$\max_{w} \mathbb{E}_w[m(pₜ)] - \frac{1}{\beta_c} \mathrm{KL}(w \,\|\, \text{Uniform})$$

produces the exponential family weight $w^*(pₜ) \propto \exp[\beta_c(\omega - pₜ)]$, ensuring smooth, bounded, and differentiable sample emphasis.

Dual-Phase Class Coupling

In the low-confidence regime ($pₜ<\omega$), $f'_c=f_c$ so that tail classes ($f_c\ll1$) yield $e-f_c\approx e$, resulting in strong sample weights. For high-confidence ($pₜ\ge\omega$), $f'_c=1-f_c$ so that head classes ($f_c\approx1$) result in $e-(1-f_c)\to e$, but the exponent is negative, thus suppressing their influence. This mechanism ensures adaptive reweighting tuned to both per-sample difficulty and class rarity.

Gradient Modulation

Differentiating the loss with respect to logits shows that:

$$\nabla_z \mathcal{L}_{\text{total}} = \Psi(pₜ,\gamma_c)\cdot(p-e_t)$$

where $\gamma_c = e - f'_c(pₜ)$ and:

$$\Psi(pₜ,\gamma_c)=\gamma_c^{\omega-pₜ}\cdot[1-pₜ\cdot\ln\gamma_c\cdot\ln pₜ]$$

This imbues two properties:

• Boundedness: $x\ln x$ is bounded on $[0,1]$, so $\Psi$ remains finite as $pₜ\to0$.
• Entropy awareness: $-pₜ\ln pₜ$ prioritizes uncertain regions.

3. Integration into Loss Frameworks

Ω enters the learning objective as a multiplicative factor applied to any base loss $\mathcal{L}_{\text{base}}(pₜ)$—including standard cross-entropy or its adjusted variants. For a batch of size $N$:

$$L = \frac{1}{N}\sum_{i=1}^N \Omega(pₜ^{(i)}, f_{y_i}) \cdot [-\ln pₜ^{(i)}]$$

This formulation directly modulates per-sample gradient magnitudes and does not alter inference-time logits. No normalization constant is applied since Ω operates only as a reweighting during training.

4. Implementation and Stability Properties

• Probability $pₜ$ estimation: Standard softmax over the model’s logit vector $z\in\mathbb{R}^K$.
• Class frequency $f_c$ computation: Precomputed as $N_c/N$ from the training set and treated as fixed.
• Hyperparameter selection: Pivot $\omega$ is selected via validation, with $\omega=0.75$ yielding near-optimal tradeoffs.
• Numerical guarantees: $pₜ$ is strictly positive, and $\ln(e-f'_c)$ remains well-behaved since $f'_c\in[0,1]$.
• Code integration: In PyTorch, Ω is implemented as a single-line multiplier atop the standard loss, e.g.,

  weight = (math.e - f_prime)**(omega - p_t) loss = weight * base_loss

5. Theoretical Guarantees

• Continuity: Ω is continuous at $pₜ=\omega$, with $\Omega(\omega, f_c)=1$ and one-sided derivatives bounded by $\ln(e/(e-1))\approx0.46$, precluding pathological kinks.
• Lipschitz behavior: The transition at the pivot is bounded, yielding stable SGD updates.
• Gradient boundedness and entropy scaling: The design prevents gradient explosion—particularly relevant for rare-tail samples—and emphasizes training on uncertain (high-entropy) predictions.

6. Empirical Evaluation

Empirical validation across three long-tailed benchmarks demonstrates that CCAR’s Ω(pₜ, f_c) consistently enhances tail-class performance while maintaining or improving head-class accuracy.

Dataset	Baseline / Method	Top-1 Accuracy	+CCAR Ω Improvement
CIFAR-100-LT (IF=200)	CE / Ours+CE	34.84 → 36.12	+1.28
	BS / Ours+BS	43.30 → 46.20	+2.90
CIFAR-100-LT (IF=50)	LDAM-DRW / Ours+LA	47.62 vs. 52.78	+5.16
ImageNet-LT	CE / Ours+CE	41.60 → 45.62	+4.02
	Logit Adjust / Ours+LA	51.10 → 52.76	+1.66
iNaturalist2018	CE / Ours+CE	61.70 → 64.90	+3.20
	BS / Ours+BS	69.80 → 70.10	+0.30

(A full breakdown appears in (Jagati et al., 22 Jan 2026), including ablation for ω and stratified accuracy for medium/few-class splits.)

Ablation studies reveal that $\omega=0.75$ balances the amplification for uncertain samples and suppression for over-confident predictions across both large-scale and smaller benchmarks. Figure references from the original work (gradient modulation analysis and the 3D Ω surface) illustrate the smooth, two-phase scaling behavior with respect to class frequency and confidence.

7. Significance and Outlook

Ω(pₜ, f_c) is a lightweight, theoretically grounded, and plug-and-play solution for long-tailed class imbalance. Its dual-phase structure offers an interpretable mechanism for boosting rare-class influence when needed and self-suppressing overconfident majority-class samples, all while maintaining stability and computational efficiency. The function complements logit-adjusted and margin-based losses, requires no additional post-hoc normalization, and integrates seamlessly into modern deep learning frameworks. Results across diverse datasets indicate robust improvements, particularly for underrepresented classes in imbalanced settings (Jagati et al., 22 Jan 2026). A plausible implication is that future class- and confidence-adaptive reweighting schemes may revisit the dual-phase mechanism instantiated by Ω for even broader application domains.