![[2602.17853] Neural Prior Estimation: Learning Class Priors from Latent Representations](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.17853] Neural Prior Estimation: Learning Class Priors from Latent Representations
Summary
The paper introduces Neural Prior Estimator (NPE), a framework for learning class priors from latent representations, addressing class imbalance in deep neural networks.
Why It Matters
Class imbalance is a significant challenge in machine learning, leading to biased predictions. NPE provides a theoretically grounded solution that enhances model performance, especially for underrepresented classes, making it relevant for researchers and practitioners in AI and machine learning.
Key Takeaways
- NPE learns feature-conditioned log-prior estimates, improving bias-aware predictions.
- The framework operates without needing explicit class counts or hyperparameters.
- Experiments show consistent performance improvements on long-tailed datasets.
Computer Science > Machine Learning arXiv:2602.17853 (cs) [Submitted on 19 Feb 2026] Title:Neural Prior Estimation: Learning Class Priors from Latent Representations Authors:Masoud Yavari, Payman Moallem View a PDF of the paper titled Neural Prior Estimation: Learning Class Priors from Latent Representations, by Masoud Yavari and Payman Moallem View PDF HTML (experimental) Abstract:Class imbalance induces systematic bias in deep neural networks by imposing a skewed effective class prior. This work introduces the Neural Prior Estimator (NPE), a framework that learns feature-conditioned log-prior estimates from latent representations. NPE employs one or more Prior Estimation Modules trained jointly with the backbone via a one-way logistic loss. Under the Neural Collapse regime, NPE is analytically shown to recover the class log-prior up to an additive constant, providing a theoretically grounded adaptive signal without requiring explicit class counts or distribution-specific hyperparameters. The learned estimate is incorporated into logit adjustment, forming NPE-LA, a principled mechanism for bias-aware prediction. Experiments on long-tailed CIFAR and imbalanced semantic segmentation benchmarks (STARE, ADE20K) demonstrate consistent improvements, particularly for underrepresented classes. NPE thus offers a lightweight and theoretically justified approach to learned prior estimation and imbalance-aware prediction. Subjects: Machine Learning (cs.LG); Computer Vision and Patter...