---
title: 'CHEST Loss: Combining Hyperbolic & Euclidean Metrics'
url: https://www.emergentmind.com/topics/chest-loss
type: topic
---

# CHEST Loss: Combining Hyperbolic & Euclidean Metrics

Searching arXiv for the CHEST loss paper and closely related metric-learning work.
CHEST loss, short for **Combined Hyperbolic and Euclidean Soft Triple loss**, is a deep metric learning objective that combines **proxy-based losses in hyperbolic and Euclidean spaces** with a **regularization loss based on hyperbolic hierarchical clustering**. It was introduced to address a specific gap in deep metric learning: hyperbolic-space methods had been based on pair-based loss or unsupervised regularization loss, whereas supervised proxy-based losses in hyperbolic space had not been reported yet because of difficulties in applying proxy-based losses in a hyperbolic space. The method targets large-scale datasets, where proxy-based losses are attractive since they have less training complexity, and it was evaluated on four benchmark datasets, achieving a new state-of-the-art performance [2510.05643].

## 1. Motivation and problem setting

Deep metric learning seeks an embedding space in which semantic similarity is reflected by geometric proximity. Within that setting, CHEST loss is motivated by the complementary strengths and weaknesses of Euclidean and hyperbolic geometry. Euclidean space is traditionally used in deep metric learning and is stable and well understood, but it is less capable of modeling data with inherent hierarchy. Hyperbolic space is attractive for deep metric learning since it can represent richer structures, such as tree structures, but direct proxy learning in hyperbolic space presents stability and scale challenges [2510.05643].

The stated intuition is that **hyperbolic distance scales much more aggressively**, and gradients become unstable near the boundary of the Poincaré ball, making direct proxy learning difficult. By contrast, **Euclidean proxy losses such as SoftTriple are stable** but may miss hierarchical relations in the data. CHEST loss therefore couples both spaces in a single objective: the Euclidean component contributes stable optimization, while the hyperbolic component contributes hierarchical expressivity. In the formulation given for CHEST loss, each loss acts as a regularizer for the other, improving generalization and stability [2510.05643].

## 2. Geometric construction and proxy parameterization

CHEST loss is defined over two coupled embedding spaces. In the Poincaré ball model, the hyperbolic space with curvature parameter \(c\) is written as
\[
\mathbb{D}_c^n = \left\{ \mathbf{x} \in \mathbb{R}^n \;\middle|\; c\|\mathbf{x}\|^2 < 1 \right\}.
\]
The method uses data embeddings in Euclidean space and in hyperbolic space, with the hyperbolic representation obtained via an exponential map. The proxy set is defined in Euclidean space and then mapped into hyperbolic space, so the two branches are strongly coupled through shared proxy parameters [2510.05643].

This design is central to the method. Proxies are learnable class representatives; CHEST loss uses **multiple proxies per class**, in the SoftTriple style, rather than a single proxy. Euclidean-space similarity and hyperbolic-space similarity are each computed from distances to these class-specific proxies, and both similarities enter the final supervised objective. A plausible implication is that the method retains the optimization behavior of proxy-based metric learning while expanding the representational capacity available to the embedding through the hyperbolic branch.

The method also introduces a distinct regularization term that is only applied in hyperbolic space. This regularizer is based on **hyperbolic hierarchical clustering**, and it exploits the stated ability of hyperbolic geometry to reflect tree-like structure. In that sense, the Euclidean branch supplies stable proxy supervision, whereas the hyperbolic branch supplies both proxy supervision and hierarchy-sensitive organization [2510.05643].

## 3. Objective function and constituent losses

The primary supervised component is a weighted sum of SoftTriple-like losses in the two spaces:
\[
\mathcal{L}_{sim}(\mathbf{x}_i) = \eta_H \mathcal{L}_H(\mathbf{x}_i) + \eta_E \mathcal{L}_E(\mathbf{x}_i),
\]
where \(\eta_H\) and \(\eta_E\) are weighting coefficients for the hyperbolic and Euclidean terms. The formulation given for each branch uses class similarities, scale hyperparameters \(\lambda_*\), and margin parameters \(\delta_*\), with \(*\) denoting either the hyperbolic or Euclidean space [2510.05643].

The hyperbolic hierarchical clustering regularizer is defined through hyperbolic proxy similarities
\[
\mathcal{S}_{i,j} = \exp \left(- D_H(\mathbf{p}_i, \mathbf{p}_j)\right),
\]
and is evaluated on triplets of proxies consisting of an anchor, a same-class proxy, and a different-class proxy. The paper denotes this regularizer as \(\mathcal{L}_{\mathrm{HypHC}}\) and uses it to impose a hierarchical structure in hyperbolic space by regularizing the arrangement of proxies [2510.05643].

The full CHEST loss is
\[
\mathcal{L} = \frac{1}{N} \sum_{i=1}^{N} \left( \eta_H \mathcal{L}_H(\mathbf{x}_i) + \eta_E \mathcal{L}_E(\mathbf{x}_i) \right) + \frac{\tau}{M} \sum_{i=1}^{M} \mathcal{L}_{\mathrm{HypHC}}(t_i),
\]
where \(\tau\) controls the contribution of the regularization term. This expression makes the method structurally explicit: CHEST loss is not a single-space loss with an auxiliary penalty, but a composite objective in which Euclidean proxy supervision, hyperbolic proxy supervision, and hyperbolic hierarchical regularization are optimized jointly [2510.05643].

## 4. Optimization behavior, regularization, and ablation findings

The optimization rationale of CHEST loss follows directly from the geometry of the two spaces. The data state that **each loss regularizes the other**: if learning is unstable in hyperbolic space, the Euclidean loss provides good gradients, and conversely the coupled training improves both stability and generalization. This is a stronger claim than merely saying that two branches are ensembled at inference time; the interaction is part of the training objective itself [2510.05643].

The reported ablations support that interpretation. Removing either the hyperbolic or the Euclidean loss degrades performance, showing that both are necessary. Using only the hyperbolic loss leads to overfitting, with Recall@1 decreasing in late training. The addition of hierarchical regularization and multiple proxies further improves performance, but the primary gain comes from combining both spaces. The paper also reports histograms comparing similarity distributions, showing that the combined loss provides better separation between positive and negative classes [2510.05643].

A plausible implication is that CHEST loss should be understood as a **coupled geometric regularization scheme** rather than merely a hybrid metric. The Euclidean term stabilizes optimization and proxy learning; the hyperbolic term and \(\mathcal{L}_{\mathrm{HypHC}}\) shape the embedding toward hierarchical organization. The result is a metric-learning objective that addresses both learning stability and structural expressivity within a single training criterion.

## 5. Empirical performance and benchmark results

CHEST loss was evaluated on **CUB200**, **Cars196**, **In-shop**, and **Stanford Online Products (SOP)**. The paper compares against leading deep metric learning methods including **SoftTriple**, **MPA**, **Hyp-ViT**, **HIER**, **VPTSP-G**, and **RS@K**, and reports that CHEST loss achieves the highest Recall@1 values on all four datasets, surpassing previous state of the art or tying it on SOP [2510.05643].

| Dataset | Recall@1 with CHEST-H | Previous best |
|---|---:|---:|
| CUB200 | **88.8** | 88.5 |
| Cars196 | **93.1** | 91.2 |
| In-shop | **94.5** | 92.5 |
| SOP | **88.0** | 88.0 |

The benchmark evidence is complemented by qualitative analysis. In retrieval visualizations, the embeddings are reported to robustly capture semantic similarity, even across varying poses and backgrounds, and the combination of the two spaces allows the model to retrieve more semantically similar results than a single-space alternative. The paper also states that both hyperbolic and Euclidean output spaces see performance boosts, which it interprets as evidence of mutual regularization [2510.05643].

## 6. Scope, applications, and terminological distinctions

The applications explicitly identified for CHEST loss are **large-scale metric learning tasks** where training complexity and generalization matter, including **image retrieval**, **few-shot learning**, and **person re-identification**. The method is especially suited to data with **hierarchical or tree-like structure**, such as **biological taxonomies**, **product catalogs**, or **semantic image grouping**, and to settings where training stability is crucial [2510.05643].

A recurrent source of confusion is the visual similarity between the acronym **CHEST** and the broader literature on **chest** imaging. In the material summarized here, the term **CHEST loss** refers specifically to **Combined Hyperbolic and Euclidean Soft Triple Loss** in deep metric learning [2510.05643]. By contrast, chest X-ray papers discuss different loss designs. For example, long-tailed multi-label chest X-ray classification emphasizes **LDAM-DRW** rather than CHEST loss [2603.02294], and the paper **"Multi-loss ensemble deep learning for chest X-ray classification"** states that it does **not use the term "CHEST Loss" per se** [2109.14433]. Likewise, the **CheX-Nomaly** paper states that its customized sigmoid focal cross-entropy loss with a transform step is **not named "CHEST Loss"** [2311.01777].

This suggests a useful terminological distinction. In current usage within the material provided, **CHEST loss** denotes a specific proxy-based deep metric learning objective that couples hyperbolic and Euclidean spaces, rather than a generic label for loss functions used in chest radiography.

Source: https://www.emergentmind.com/topics/chest-loss