Tahar Chettaoui, Guray Ozgur, Eduarda Caldeira, Naser Damer, Fadi Boutros
8 min
Abstract
Multi-head attention in Vision Transformers (ViTs) allows independent attention patterns across heads, but without explicit regularization, heads may learn redundant patterns rather than complementary roles. Correlation-based penalties encourage head disentanglement, yet apply uniform regularization across all training samples. In face recognition (FR), hard samples whose embeddings lie close to competing identities are critical for optimization and generalization. Motivated by this, we propose ViT-HarDis, which conditions attention head disentanglement on sample difficulty: samples with high cosine similarity to negative class centers receive stronger head decorrelation, enforcing specialization where identity ambiguity is highest. Experiments show improvements over both the baseline and uniform disentanglement on most benchmarks. Our analysis shows that ViT-HarDis reduces the similarity of hard samples to competing class centers while decreasing inter-head attention similarity, and that applying disentanglement only to the final transformer layers outperforms regularizing the entire network. Trained on WebFace4M, ViT-HarDis achieves SOTA results among ViT-based methods on IJB-B, IJB-C, and TinyFace Rank-1, while achieving the best IJB-B and TinyFace results when trained on MS1MV3, improving FR without architectural modifications or additional inference cost. Code: https://github.com/TaharChettaoui/Hard-Sample-Conditioned-Attention-for-FR.
Alex: So the important case is a face assigned to the correct identity, but still uncomfortably close to someone else.
Sam: Yes. The method measures similarity to the closest competing center. Greater similarity means a stronger penalty when attention heads overlap. The ordinary face-recognition objective still trains identity discrimination alongside that extra penalty.
Alex: Why leave out the actual mistakes? Those sound like the hardest examples of all.
Sam: The authors exclude currently misclassified samples from this extra penalty to avoid amplifying possible annotation errors. They are not removed from face-recognition training. And exclusion is temporary: a sample becomes eligible when its own identity center becomes the closest.
Alex: That exclusion could itself explain the improvement. Did they separate its effect from the difficulty weighting?
Sam: They tested a variant that also weights misclassified samples. It still beats uniform regularization on the large-scale benchmarks and TinyFace, their low-resolution evaluation. Excluding mistakes helps on most comparisons, but not all. So the evidence supports difficulty weighting as more than a side effect of exclusion.
Alex: How do they avoid simply increasing the total regularization and calling that adaptation?
Sam: They normalize the weights within each training batch, keeping their average approximately constant. That redistributes the pressure rather than deliberately scaling it up. The realized penalty can still differ because sample weights and attention overlap may be related.
Alex: And can the model manipulate its difficulty score to escape the penalty?
Sam: The difficulty weights are held fixed during each gradient calculation. This extra loss changes the model through its attention maps, not through the calculation of difficulty. That keeps the score acting as a weighting signal.
Alex: Give us a concrete picture of what specialization looks like in their analysis.
Sam: They examine the hardest tenth of training faces, selected using the baseline model. With the baseline and uniform penalty, many heads focus on similar regions, such as both eyes. With ViT-HarDis, the averaged maps show more distinct regional patterns.
Alex: Those visualizations help, but they do not establish better identity separation. Is there a separate measurement?
Sam: For those selected hard faces, average similarity to the closest competing identity falls from point four with the baseline to point three one with ViT-HarDis. Uniform regularization lands between them. That supports reduced ambiguity in the learned representations, alongside lower similarity between heads.
Alex: Let’s move from that internal behavior to recognition performance. What is the cleanest comparison?
Sam: The main ablations use a smaller transformer trained on approximately six hundred thousand images from an identity subset. They compare the unchanged baseline, uniform regularization, and difficulty-weighted regularization. The backbone includes landmark-based position information, so this is a specific, strong baseline rather than an arbitrary transformer.
Alex: And how much does the difficulty weighting buy on a challenging benchmark?
Sam: On IJB-B, a large-scale verification benchmark, it improves true acceptance over baseline by about one point three percentage points. That comparison holds the false-acceptance rate fixed. Uniform regularization produces a smaller gain at the same operating point.
Alex: A system’s operating threshold matters. Does the improvement survive when false matches are restricted more tightly?
Sam: Not everywhere. At the stricter reported IJB-B threshold, the main proposed version is slightly below baseline, though above uniform regularization. On IJB-C, the other large-scale verification benchmark, it improves at both reported thresholds. “Most benchmarks” is the defensible summary, not “every operating condition.”
Alex: There is also a question of where to apply the penalty. Does every layer benefit from being pushed apart?
Sam: Their layer-selection experiment says no. Regularizing only the final quarter of transformer layers gives the best average and large-scale benchmark results. Applying it throughout the network performs worst among the tested selections. Stronger specialization everywhere is not the lesson.
Alex: Do the findings extend beyond that reduced training set?
Sam: They also train larger models on full datasets, including WebFace4M, which contains four million images. With that dataset, they report the highest results among the compared transformer methods on IJB-B, IJB-C, and TinyFace’s top-match accuracy. They also retrain their baseline under matching settings for a fairer comparison.
Alex: What should keep us from treating those rankings as a general solution to robust recognition?
Sam: The comparisons concern transformer-based face recognition, not every model family or application. Detailed ablations use the reduced dataset, and competing published systems differ in configuration. Full-scale gains are modest and benchmark-dependent. The evidence supports this intervention in the tested setting, not a universal recipe.
Alex: And the training bill is real, even though deployment stays unchanged.
Sam: On their reported WebFace4M setup, training takes roughly thirty percent longer. No parameters or operations are added at inference. So the trade is more training computation for gains without a more expensive deployed model.
Alex: Who should read the full document, and where would you send them first?
Sam: Researchers training transformer face-recognition models should start with Hard Sample-conditioned Disentanglement, especially the comparison against uniform weighting. Then read Layer Selection for Attention Disentanglement before copying the loss. Check the full-data comparisons against the retrained baseline, not just the headline rankings.
Alex: And for researchers outside face recognition, what is the idea to carry away?
Sam: Useful specialization is not just about making components different. It is about choosing which examples, and which layers, need that difference.
Alex: A reason to ask where diversity helps, not just whether you can enforce it.