Proxy-Decidability Loss (PD-Loss) Explained

• Proxy-Decidability Loss (PD-Loss) is a deep metric learning objective that estimates genuine and impostor score distributions via learnable class proxies instead of exhaustive pairwise comparisons.
• It leverages a statistical separability measure derived from the decidability index to optimize embedding similarity, thereby improving verification performance.
• PD-Loss overcomes traditional pairwise loss limitations by combining proxy-based methods with a global distribution-aware objective, enhancing both scalability and convergence.

Searching arXiv for the primary PD-Loss and precursor D-Loss papers. Proxy-Decidability Loss (PD-Loss) is a deep metric learning objective that combines the distribution-level separability criterion of Decidability Loss (D-Loss) with the computational structure of proxy-based learning. It estimates the two score distributions central to verification-style embedding learning—genuine and impostor—through learnable class proxies rather than through exhaustive sample-sample comparisons, and then optimizes their separability via a log-form objective derived from the decidability index $d'$ (Silva et al., 23 Aug 2025). In this formulation, PD-Loss is positioned as a proxy-based estimator of a statistical objective: it retains the emphasis on global distribution structure inherited from D-Loss while addressing the large-mini-batch and pairwise-cost limitations of the original decidability-based approach (Silva et al., 2021).

1. Genealogy and problem setting

PD-Loss emerges at the intersection of two established strands in deep metric learning. The first strand consists of pairwise and triplet-style objectives, including contrastive loss, triplet loss, N-pair, Multi-Similarity (MS), and Circle loss. These methods directly enforce local relations among positives and negatives, but they are described as requiring careful mining of informative or hard pairs, exhibiting strong dependence on batch composition, typically needing margin or scale hyperparameters, and often converging slowly because they use only a subset of sample relationships in a batch (Silva et al., 23 Aug 2025). The second strand consists of proxy-based methods such as ProxyNCA and ProxyAnchor, which replace sample-sample comparisons with sample-proxy comparisons and thereby improve scalability and convergence, but do not explicitly optimize the separability of the full genuine and impostor distributions in the embedding space (Silva et al., 23 Aug 2025).

The immediate precursor of PD-Loss is D-Loss, introduced as a loss function for embedding learning and verification rather than classification (Silva et al., 2021). D-Loss was motivated by the observation that Softmax / cross-entropy, while strong for classification, does not explicitly optimize verification geometry, whereas triplet losses are better aligned with verification but depend on anchor, positive, and negative selection, often require hard mining, need careful margin tuning, and can converge slowly (Silva et al., 2021). D-Loss answered this by directly optimizing the decidability index computed from batchwise genuine and impostor pairwise distances. PD-Loss preserves that distribution-aware target but changes the estimation mechanism: genuine distribution is approximated by sample similarity to the correct class proxy, and impostor distribution by sample similarity to incorrect class proxies (Silva et al., 23 Aug 2025).

This placement is central to understanding the term “Proxy-Decidability.” The “proxy” component refers to class representatives learned jointly with the embedding network; the “decidability” component refers to the biometric notion of distribution separability. A plausible implication is that PD-Loss should be interpreted less as a conventional classification surrogate and more as a statistical metric-learning objective whose sufficient statistics are approximated through proxies.

2. Decidability as the organizing principle

The conceptual basis of both D-Loss and PD-Loss is the decidability index $d'$, a biometric statistic that measures how well two score distributions are separated. If $\mu_g, \sigma_g^2$ denote the mean and variance of the genuine distribution and $\mu_i, \sigma_i^2$ denote the mean and variance of the impostor distribution, then

$$d' = \frac{|\mu_i - \mu_g|}{\sqrt{(\sigma_g^2 + \sigma_i^2)/2}}$$

as stated in the PD-Loss formulation (Silva et al., 23 Aug 2025). The D-Loss paper gives the same statistic in the equivalent notation $\mu_G, \mu_I, \sigma_G^2, \sigma_I^2$ and uses it as the core criterion for embedding quality in verification (Silva et al., 2021).

The statistical interpretation is direct. Larger separation between the means of genuine and impostor scores increases $d'$, while larger variances decrease it by increasing overlap between the two distributions. In D-Loss, these scores are Euclidean distances between embeddings, so good representations should yield small genuine distances and large impostor distances (Silva et al., 2021). In PD-Loss, the scores are similarities between normalized embeddings and normalized proxies, so the direction is inverted: good representations should yield high genuine similarity and low impostor similarity (Silva et al., 23 Aug 2025).

This shift from distance to similarity is not merely cosmetic. It changes the sign structure of the optimization while preserving the statistical target. The D-Loss paper minimizes the inverse of decidability,

[ \mathcal{L}_{D\mbox{-}loss