Source-linked AI summary
Ring loss: Convex Feature Normalization for Face Recognition
Yutong Zheng, Dipan K. Pal, Marios Savvides
TL;DR
Face recognition requires robust classification with many classes, few samples per class, and nuisance transformations, while direct feature normalization is non-convex. The paper proposes Ring loss, a convex, differentiable augmentation that softly constrains features toward a learned norm. Ring loss provides consistent improvements across face-recognition baselines and datasets, including state-of-the-art results on Janus CS3 and robustness at low resolution.
Problem
Face recognition combines many classes, few samples per class, and nuisance transformations, while direct feature normalization yields a non-convex formulation.
Method
Ring loss adds a convex, differentiable norm constraint to a primary loss and learns the target feature norm during training.
Results
Ring loss consistently improves Softmax and SphereFace across evaluations, achieving state-of-the-art performance on challenging Janus CS3.
Takeaways & Limitations
Feature normalization can balance angular classification margins, align training with cosine-based testing, and improve robustness to lower-resolution faces.
Takeaways & Limitations
For α = 50, gradients exploded in the relatively deep Res64 architecture, while learning α initialized at 30 did not converge.
Abstract
from arXiv · showhide
We motivate and present Ring loss, a simple and elegant feature normalization approach for deep networks designed to augment standard loss functions such as Softmax. We argue that deep feature normalization is an important aspect of supervised classification problems where we require the model to represent each class in a multi-class problem equally well. The direct approach to feature normalization through the hard normalization operation results in a non-convex formulation. Instead, Ring loss applies soft normalization, where it gradually learns to constrain the norm to the scaled unit circle while preserving convexity leading to more robust features. We apply Ring loss to large-scale face recognition problems and present results on LFW, the challenging protocols of IJB-A Janus, Janus CS3 (a superset of IJB-A Janus), Celebrity Frontal-Profile (CFP) and MegaFace with 1 million distractors. Ring loss outperforms strong baselines, matches state-of-the-art performance on IJB-A Janus and outperforms all other results on the challenging Janus CS3 thereby achieving state-of-the-art. We also outperform strong baselines in handling extremely low resolution face matching.
1. Introduction
Face recognition exposes challenges for supervised classification, including many classes, few samples per class, nuisance transformations, and unresolved questions about feature normalization. Ring loss addresses normalization by adding a convex, learnable-norm augmentation to standard losses such as Softmax.
- Face recognition is difficult because it involves many classes, relatively few training samples per class, and significant nuisance transformations.
- 99.34% accuracy with Ring loss versus 98.97% with Softmax demonstrates improved MNIST performance from learned normalized features.
- Feature normalization can improve performance but direct normalization in deep networks produces a non-convex formulation.
- Ring loss normalizes sample features through a convex augmentation of a primary loss while learning the target norm during training.
- Feature matching commonly uses cosine distance at test time, while training protocols may omit normalization; Ring loss eliminates this disconnect.
2. Ring loss: Convex Feature Normalization
Softmax can produce unequal feature norms, creating imbalanced angular margins and a mismatch with cosine-based testing. Ring loss replaces hard normalization with a convex, learned soft constraint that guides features toward a common norm.
- Ring loss: Ring loss reduces norm imbalance and supports more balanced angular margins while integrating differentiably with Softmax or other loss functions.The method is intended to align training with cosine-based feature matching and mitigate errors associated with low-norm features.
- Motivation: Softmax assigns different losses to same-direction features when their norms differ, producing unequal angular classification margins across classes.Classes with different feature norms receive different effective angular margins during training.
- Motivation: Low-norm features exhibit greater angular variation, which can increase cosine-metric test errors and widen the training–testing mismatch.The mismatch arises because training commonly uses Softmax while face-recognition testing uses cosine distance.
- Motivation: Hard feature normalization imposes a norm-equality constraint whose feasible set is non-convex, so standard SGD does not ensure that the network learns normalized outputs.Relaxing equality to an inequality restores convexity but does not necessarily enforce equal norms.
- Ring loss: Ring loss adds a squared penalty between each feature norm and a learned target norm, with λ controlling its trade-off against the primary loss.The squared difference produces larger updates when a sample norm is far from the target, supporting faster convergence.
- Ring loss: In simulations, λ = 0 fails to converge, λ = 10 produces very small Softmax-directed steps, and λ = 1 achieves a good balance.The simulations update mesh points for 20 steps with target norm R = 1; large-scale experiments found a broad range of λ values effective.
3. Experimental Validation
Ring loss is evaluated across large-scale face-recognition benchmarks, where it consistently improves Softmax and SphereFace baselines, including difficult pose, scale, and resolution conditions.
- Benchmark setup: Ring loss is benchmarked with Softmax and SphereFace on LFW, IJB-A Janus, Janus CS3, CFP, and MegaFace.The experiments also test low-resolution face matching on downsampled Janus CS3 images.
- LFW: 99.52% LFW accuracy is achieved by Softmax with Ring loss, compared with 98.47% for Softmax alone.SphereFace changes from 99.47% to 99.50%; the authors note larger protocols show clearer trends than LFW’s 6,000 matches.
- IJB-A Janus: 78.41% verification at 10^-5 FAR is reported for Softmax with Ring loss on IJB-A Janus, versus 60.52% for Softmax.Ring loss also raises SphereFace from 78.52% to 82.41%, matching the cited state-of-the-art despite using a shallower architecture.
- Janus CS3: 82.74% verification at 10^-4 FAR is achieved by Ring-augmented SphereFace on Janus CS3, outperforming the other reported models.At 10^-6 FAR, the same model reaches 35.18%, reported as state-of-the-art on this challenging benchmark.
- MegaFace: 71.67% MegaFace identification is obtained by Softmax with Ring loss, up from 56.36% for Softmax alone against 1 million distractors.SphereFace improves from 74.95% to 75.22%, while Ring loss is described as easier to tune than l2-constrained Softmax.
- CFP and robustness: 87.43% CFP verification is reported for Ring-augmented Softmax versus 55.86% for Softmax, demonstrating improved frontal-to-profile matching.SphereFace reaches 90.94% with Ring loss at λ = 0.01, while l2-constrained Softmax reaches 83.69%.