Max-Severity Triplet Loss

• Max-Severity Triplet Loss is a metric-learning framework that defines severity as the violation magnitude of task-specific constraints, especially in weakly supervised settings.
• It realigns over-severe target predictions by using patient-level max severity labels to pull features toward admissible source prototypes.
• The method, integrated with Shared Aggregation Tokens and a ResNet-18 backbone, shows improved domain adaptation performance compared to other triplet mining strategies.

Searching arXiv for papers on “Max-Severity Triplet Loss” and closely related triplet-hardness formulations. Max-Severity Triplet Loss denotes a triplet-based metric-learning construction in which “severity” is not merely a synonym for sample hardness, but is tied to the most severe permissible relation under a task-specific constraint. The term appears explicitly in weakly supervised domain adaptation for ulcerative colitis severity estimation, where patient-level diagnoses are determined by the most severe region within a patient, and the loss is used to correct target-domain instances whose predicted severity exceeds the bag-level maximum (Yamaguchi et al., 18 Sep 2025). Related triplet-learning literature uses closely adjacent ideas—batch-local maximum violation, hardest-positive/hardest-negative mining, and extreme-margin regimes—in ways that clarify what “max severity” can mean operationally, but these usages are not identical (Hermans et al., 2017, Martzloff et al., 2023).

1. Terminological scope and relation to triplet hardness

In classical metric learning, the standard triplet loss enforces

$$D(f_\theta(x_a), f_\theta(x_p)) + m \le D(f_\theta(x_a), f_\theta(x_n)),$$

with anchor $x_a$, positive $x_p$, and negative $x_n$, and penalizes violations through

$$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$

This makes “severity” naturally interpretable as the magnitude of the violation term $m + D_{a,p} - D_{a,n}$ (Hermans et al., 2017).

The literature associates this severity with three distinct mechanisms. In person re-identification, batch-hard triplet training selects, for each anchor, the farthest positive and nearest negative within a structured mini-batch, which is a batch-local maximum-violation rule (Hermans et al., 2017). In weakly supervised domain adaptation for ulcerative colitis, Max-Severity Triplet Loss uses patient-level max labels to identify clinically impossible target instances and re-align them to source prototypes of admissible severity (Yamaguchi et al., 18 Sep 2025). In fairness-oriented work on triplet collapse, “max severity” is pushed to an extreme by setting the margin beyond the latent-space diameter so that all triplets are active simultaneously (Martzloff et al., 2023). This suggests that the phrase is best understood as a family of severity-maximizing or severity-constrained triplet formulations rather than a single universal loss.

Context	Mechanism	Severity interpretation
Person re-identification	Batch-hard triplet	Hardest positive, hardest negative within batch
Weakly supervised UC adaptation	Max-Severity Triplet Loss	Instance exceeds bag-level max severity
Fairness via collapse	Outrageous-margin triplet	All triplets forced active

2. Exact formulation in ulcerative colitis severity estimation

In the ulcerative colitis setting, source bags are

$$\mathcal{B}_i^\mathrm{s} = \{\mathbf{x}_{i,j}^\mathrm{s}, y_{i,j}^\mathrm{s}\}_{j=1}^{|\mathcal{B}_i^\mathrm{s}|},$$

target bags are

$$\mathcal{B}_i^\mathrm{t} = \{\mathbf{x}_{i,j}^\mathrm{t}\}_{j=1}^{|\mathcal{B}_i^\mathrm{t}|},$$

instance labels in the source are $y_{i,j}^\mathrm{s} \in \{1,2,\ldots,K\}$, and bag labels in both domains are $$Y_i^\mathrm{s}, Y_i^\mathrm{t} \in \{1,2,\ldots,K\}$$ with ordinal order $x_a$0 (Yamaguchi et al., 18 Sep 2025). The defining clinical constraint is the max-severity bag property

$x_a$1

Source and target instance features are

$x_a$2

and target predicted instance labels are denoted $x_a$3 (Yamaguchi et al., 18 Sep 2025). For each source severity class $x_a$4, the source prototype is the mean feature

$x_a$5

The Max-Severity Triplet Loss is applied only to target instances that violate the bag-level maximum, namely those satisfying

$x_a$6

For a target bag $x_a$7, a positive prototype $x_a$8 is chosen from an allowed severity class, and a negative prototype $x_a$9 is chosen from a higher-severity class. Using Euclidean distance and margin $x_p$0, the loss is

$x_p$1

subject to

$x_p$2

The hinge condition enforces

$x_p$3

Operationally, the anchor is a target instance whose predicted severity is too high for its patient, the positive is a source prototype consistent with the bag maximum, and the negative is a source prototype from a clinically impossible higher-severity class (Yamaguchi et al., 18 Sep 2025).

3. Interaction with Shared Aggregation Tokens and target-domain adaptation

The Max-Severity Triplet Loss is one part of a two-level alignment strategy. The same work introduces Shared Aggregation Tokens for global or coarse class-wise alignment at the bag level, while the triplet loss performs fine-grained instance-level correction (Yamaguchi et al., 18 Sep 2025). In the source domain, aggregation tokens

$x_p$4

are trained to attend to severity-indicative instances and produce bag representations for max-severity classification. In the target domain, these tokens are frozen, and only the target feature extractor is updated so that target bags can be classified correctly by the source-side token structure.

This division of labor is central to the method. Shared Aggregation Tokens optimize bag-level consistency, but they do not directly repair individual target instances whose predicted severity is higher than the patient-level label permits. The Max-Severity Triplet Loss fills that gap by explicitly penalizing those over-severe target instances and pulling them toward admissible source severity prototypes while pushing them away from prototypes of more severe classes (Yamaguchi et al., 18 Sep 2025). In the paper’s formulation, the full target-domain objective is

$x_p$5

with $x_p$6.

The method is trained with a ResNet-18 backbone. In the pre-training step on the source domain, the learning rates are $x_p$7 for the feature extractor and instance classifier and $x_p$8 for the bag classifier, with 1,500 epochs, batch size 16, and early stopping on 100 patients. In the domain-adaptation step, the learning rates are $x_p$9 for the discriminator and $x_n$0 for the feature extractor, with 150 epochs and batch size 16 (Yamaguchi et al., 18 Sep 2025). These details matter because the triplet term is not an isolated metric objective; it is integrated into a bag-supervised adversarial adaptation pipeline.

4. Relation to batch-hard mining and other maximum-violation triplet variants

The closest earlier formulation to “max severity” in the sense of maximum local violation is the batch-hard triplet loss of Hermans, Beyer, and Leibe. In a mini-batch constructed from $x_n$1 identities and $x_n$2 images per identity, the batch-hard loss selects, for each anchor, the hardest positive and hardest negative:

$x_n$3

The same paper describes this as a “max-violation / max-hardness” strategy within the batch and emphasizes that its hardness is local rather than global, making the selected triplets “moderately hard” (Hermans et al., 2017). A soft-margin version replaces the hinge by

$x_n$4

and the authors report that the soft-margin version of $x_n$5 is their best-performing loss (Hermans et al., 2017).

A related memory-based formulation is Triplet Online Instance Matching. TOIM keeps a Pooled Table $x_n$6 of identity-by-camera features and an Update Table of recent entries, then chooses the hard positive by

$x_n$7

and the hard negative by

$x_n$8

Its loss is

$x_n$9

which combines triplet-style hard mining with an OIM-style non-parametric softmax objective (Li et al., 2020). The paper states that the proposed loss outperforms the baseline methods by a maximum of 21.7%, including Softmax loss, OIM loss and Triplet loss (Li et al., 2020).

A further refinement appears in “Hard negative examples are hard, but useful,” which argues that hardest-negative mining is valuable but destabilizes standard triplet training because of hypersphere geometry and entanglement. The paper proposes the Selectively Contrastive Triplet loss

$$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$0

so that the most severe hard triplets are treated by a pure contrastive penalty on anchor-negative similarity (Xuan et al., 2020). This places a useful boundary on the concept of max severity: maximum-violation mining is often beneficial, but only if the loss geometry is compatible with the severity regime being enforced.

5. Extreme-margin severity and triplet collapse

A radically different use of maximum severity appears in “A Fair Classifier Embracing Triplet Collapse,” which studies the regime in which the triplet margin exceeds the maximum possible squared distance in a bounded latent space (Martzloff et al., 2023). The paper uses the hinge triplet loss

$$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$1

with squared Euclidean distance

$$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$2

The central condition is an “outrageous margin”

$$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$3

When this holds, every triplet violates the constraint, the gradients are never zero, and optimization cannot converge to a conventional margin-satisfied configuration (Martzloff et al., 2023). For the 3D output space considered in the paper, the maximum Euclidean distance depends on the final activation or normalization: $$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$4 for softmax, $$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$5 for sigmoid, $$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$6 for tanh, and $$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$7 for L1 or L2 normalization (Martzloff et al., 2023). Collapse is defined as an embedding space where samples are distributed into a small number of clusters whose intra-cluster distance is less than $$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$8.

This is not the same notion as batch-hard or prototype-based max severity. Here, maximum severity is produced by an impossible global constraint rather than by selecting the hardest permissible triplets. The paper uses this collapse mechanism as a fairness tool: with suitable stochastic triplet selection, sensitive attributes can be scrambled while some target information is retained (Martzloff et al., 2023). The broader implication is cautionary. Maximum severity can be constructive when it is local, structured, or logically constrained, but an unbounded or structurally infeasible severity regime can turn triplet learning into collapse dynamics rather than discriminative metric learning.

6. Empirical effects, limitations, and broader applicability

The ulcerative colitis domain-adaptation study provides the explicit empirical evidence tied to Max-Severity Triplet Loss. In its ablation study, the full method (“Adv + Agg. Token + Triplet”) improves substantially over ablations in the LIMUC → Private direction and remains competitive in the Private → LIMUC direction (Yamaguchi et al., 18 Sep 2025).

Direction	Method	Accuracy / Kappa / Macro-F1
LIMUC → Private	Ours	0.714 / 0.746 / 0.603
LIMUC → Private	w/o Triplet	0.658 / 0.703 / 0.576
LIMUC → Private	w/o Triplet AT	0.521 / 0.574 / 0.363
LIMUC → Private	w/o Adv Triplet AT	0.237 / 0.300 / 0.212
Private → LIMUC	Ours	0.706 / 0.787 / 0.594
Private → LIMUC	w/o Triplet	0.705 / 0.787 / 0.605
Private → LIMUC	w/o Triplet AT	0.604 / 0.645 / 0.529
Private → LIMUC	w/o Adv Triplet AT	0.580 / 0.397 / 0.344

The same study reports that adding Shared Aggregation Tokens already yields a large gain, while adding Max-Severity Triplet Loss further improves performance, particularly in the more challenging LIMUC → Private direction, where Accuracy increases from 0.658 to 0.714 and Macro-F1 from 0.576 to 0.603 (Yamaguchi et al., 18 Sep 2025). Qualitative visualization with PCA is described as showing much better overlap of source and target features per class for the full method than for adversarial learning only.

Several limitations are explicit. The loss assumes correct patient-level max-severity labels; if $$L_{\text{tri}(\theta)}=\sum_{\substack{a,p,n\ y_a = y_p \neq y_n}} \left[\, m + D_{a,p} - D_{a,n} \,\right]_+.$$9 is wrong, truly severe images may be pulled toward milder prototypes. Prototype quality depends on the source feature distribution. The triplet term is applied only to over-severe target predictions, not to under-estimated instances. For bags already at maximum severity $m + D_{a,p} - D_{a,n}$0, the constraint $m + D_{a,p} - D_{a,n}$1 prevents application of the triplet loss because there is no higher-severity negative class (Yamaguchi et al., 18 Sep 2025).

Beyond ulcerative colitis, the same source argues that the idea can extend to multiple-instance learning and weakly supervised domain adaptation settings where a bag label is determined by an aggregate such as a maximum. The examples given include cancer grading where a patient’s grade is determined by the most aggressive biopsy image, lesion burden scores where patient-level severity is the max lesion severity, sensor-based activity recognition where a session-level label reflects the most intense activity, and document-level labels based on the worst offending sentence or token (Yamaguchi et al., 18 Sep 2025). This suggests a general design pattern: use bag-level logical constraints to identify target-domain violations, then convert those violations into prototype-based triplets that enforce class-wise alignment without requiring full instance-level annotation.

In this broader sense, Max-Severity Triplet Loss is best understood as a constrained triplet-learning principle: severity is maximized or corrected relative to the strongest admissible relation in the task, and its success depends on whether that notion of severity is clinically meaningful, geometrically stable, and aligned with the supervision actually available.