Introduction
Disentangling specific features in high-dimensional data—such as "disease severity" in medical imaging or "object rotation" in computer vision—is a fundamental challenge in representation learning. While Variational Autoencoders (VAEs) are powerful tools for learning latent spaces, purely unsupervised disentanglement has been theoretically proven impossible without some form of inductive bias or supervision.
However, high-quality absolute labels are often prohibitively expensive or unavailable. In contrast, pairwise comparisons (e.g., "Patient A is sicker than Patient B") are easier for humans to provide but are frequently noisy due to observer error or subjective disagreement. This paper introduces the Robust Ordinal VAE (ROVAE), a framework designed to leverage these noisy comparisons for effective disentanglement while remaining resilient to labeling errors.
The ROVAE Framework
ROVAE extends the standard VAE architecture by incorporating two key structural innovations to handle ordinal supervision:
- Latent Splitting: The latent representation $z$ is partitioned into two distinct components: $u$, which is intended to capture the specific factor of interest (e.g., rotation), and $v$, which captures all remaining information necessary for high-fidelity reconstruction.
- Trustworthiness Gating ($s_{ij}$): To handle noisy labels, the model introduces a non-negative random variable $s_{ij}$ for each pair of compared images. This variable acts as a probabilistic gate. If the model's current latent representations for a pair contradict the provided ordinal label, $s_{ij}$ can be driven toward zero, effectively ignoring the "noisy" loss contribution for that pair.
The total objective function balances reconstruction accuracy, latent independence (controlled by a $\beta$ parameter), and the robust ordinal loss. By explicitly modeling the reliability of each comparison, ROVAE prevents erroneous labels from corrupting the gradient updates during training.
Experimental Evaluation
The authors evaluated ROVAE across several benchmarks and a real-world clinical dataset:
- Synthetic & Natural Images: Yale Faces (lighting), 3D Chairs & 3D Cars (azimuth), and UTKFace (age).
- Clinical COPD Dataset: Chest CT scans from patients with Chronic Obstructive Pulmonary Disease. Here, noisy labels were derived from the GOLD criteria, a standard but imperfect clinical severity scale.
Key Results and Disentanglement Quality
ROVAE demonstrates significant robustness to label noise. In experiments where 20% of comparison labels were intentionally flipped, ROVAE maintained high Mutual Information Gap (MIG) and $r^2$ scores, whereas standard ordinal VAEs experienced a sharp decline in performance.
On the 3D Chairs dataset, ROVAE achieved superior alignment between the latent variable $u$ and the true physical rotation compared to unsupervised baselines like $\beta$-VAE. Notably, in the clinical COPD study, the learned "severity" factor $u$ predicted physiological markers (such as lung gas trapping) with an $r^2$ of 0.558, outperforming both the original GOLD labels (0.398) and the vanilla Ordinal VAE (0.401). This suggests that ROVAE can uncover underlying biological truths even when guided by imperfect human heuristics.
Discussion and Limitations
The primary insight of ROVAE is its ability to treat supervision as a catalyst rather than an absolute truth. By identifying and filtering out contradictory labels, the model uses the remaining "clean" signal to align its latent space with meaningful real-world attributes.
However, the model's performance is sensitive to the choice of priors for the trustworthiness scores ($\mu_w$ and $\sigma_w$). If these are poorly tuned, the model may become overly skeptical of labels or, conversely, too susceptible to noise. Furthermore, while robust to moderate noise levels, the model's ability to extract a signal naturally degrades as the proportion of flipped labels approaches 50%.
Future research could extend this robust framework to more complex forms of supervision, such as relative similarity or multi-item rankings, further bridging the gap between fallible human intuition and precise machine representations.
Summary updated