Introduction: The Challenge of Ordinal Consistency
In computer vision, many classification tasks involve labels with a natural, intrinsic order. Examples include facial age estimation, disease progression stages, or image quality ratings. This field, known as Ordinal Regression (OR), differs from standard classification because the relationships between categories (e.g., age 20 is closer to 25 than to 60) are as important as the categories themselves.
While existing deep learning models for OR focus on optimizing the final prediction mapping, they often neglect the structure of the latent feature space. In many state-of-the-art models, the internal representations of samples do not necessarily reflect their ordinal relationships. For instance, the feature vector for a 60-year-old might be geometrically closer to that of a 20-year-old than a 55-year-old. This lack of "ordinal consistency" limits model robustness and interpretability.
Methodology: The CORE Regularizer
To address this, the authors propose CORE (Consistent Ordinal REpresentations), a "plug-and-play" regularizer designed to force the latent feature space to mirror the natural ordering of labels.
1. Ordinal Toset Distribution (OTD)
CORE treats a mini-batch of samples as a "totally ordered set" (toset). It calculates pairwise distances between labels to create a target probability distribution. Simultaneously, it computes distances between the corresponding feature embeddings. The goal is to align these two distributions.
2. Prototype-Constrained Optimization
Directly aligning features and labels is challenging due to the domain gap between images and numerical ranks. CORE formulates this as a constrained optimization problem using Convex Programming with Dual Decomposition. This approach derives an optimal feature embedding that respects an "ordinal prototype constraint," ensuring that samples with the same label cluster together while maintaining smooth transitions between adjacent ranks.
3. Seamless Integration
CORE is not a standalone architecture but a supplementary loss function ($L_{CORE}$). It can be integrated into existing deep OR models (such as SORD or POE) without altering their underlying structure, providing a significant performance boost with minimal overhead.
Experimental Results
The effectiveness of CORE was validated across eight datasets covering four distinct domains:
- Facial Age Estimation: On the MORPH dataset, CORE reduced the Mean Absolute Error (MAE) from 2.41 to 2.13 when added to the Mean-Variance baseline.
- Medical Imaging: In Diabetic Retinopathy classification, accuracy improved by up to 3.4%.
- Historical Dating & Aesthetics: Consistent improvements were observed in estimating the decade of historical photos and ranking image quality (AADB dataset).
Key Insights: The 1D Ordinal Manifold
One of the most significant findings is the impact of CORE on the geometry of the feature space. Visualization via t-SNE reveals that while standard methods produce scattered or clumpy clusters, CORE forces features onto a nearly one-dimensional ordinal manifold. This curved line represents the learned "sequence" of the data, demonstrating that the model has internalized the ordinal nature of the task.
Furthermore, Grad-CAM analysis shows that CORE-trained models focus on more semantically relevant features—such as skin texture for age estimation—rather than background noise, leading to better interpretability and grounding.
Conclusion
CORE provides a robust framework for ensuring that deep learning models respect the inherent structure of ordinal data. By aligning latent feature spaces with label-based tosets, it enhances both the accuracy and the internal logic of ordinal regression models across diverse applications.
Summary updated