---

title: >- [Paper Note] Geometric Regularization for Long-Tailed Semi-Supervised Learning via Gaussian Feature Bridges description: >- [ECCV 2026][Others][Gaussian Feature Bridge] Proposes Gaussian Bridge Consistency (GBC) and BridgeMix to build latent geometric interpolation paths between unlabeled samples and class anchors, significantly improving tail-class robustness in realistic long-tailed SSL. tags: - ECCV 2026 - Others - Semi-Supervised Learning - Long-Tailed Classification - Gaussian Feature Bridge - Geometric Regularization date: 2026-09-19 content_hash: 6a694bc35a440e37

Geometric Regularization for Long-Tailed Semi-Supervised Learning via Gaussian Feature Bridges

Conference: ECCV 2026
Paper: ECCV Official
Area: Others
Keywords: Semi-Supervised Learning, Long-Tailed Classification, Gaussian Feature Bridge, Geometric Regularization, Prototype Atlas

TL;DR

To tackle pseudo-label drift and confirmation bias in realistic long-tailed semi-supervised learning, this paper proposes Gaussian Bridge Consistency (GBC) to establish class-conditional latent interpolation paths between unlabeled samples and reliable anchors, coupled with confidence-guided BridgeMix for superior tail-class generalization.

Background & Motivation

In real-world semi-supervised learning (SSL), labeled and unlabeled data rarely conform to balanced, identical distributions. Instead, severe class imbalance and long-tailed distributions dominate, where majority (head) classes monopolize training signals while minority (tail) classes suffer from scarce supervision and noisy pseudo-labels. Standard SSL approaches based on thresholded pseudo-labeling or naive consistency regularization readily succumb to confirmation bias: pseudo-labels systematically drift toward head categories, progressively corrupting tail representations and causing catastrophic semantic drift. Prior remedies in long-tailed SSL (LTSSL) predominantly operate at the output or logit level—such as class-adaptive thresholding, logit calibration, or loss reweighting—leaving the intermediate latent representations vulnerable and unregularized.

The fundamental tension stems from the high uncertainty and variance of unlabeled tail features near noisy decision boundaries. Conventional methods lack a continuous, geometry-aware mechanism to transport these drifting features toward genuine class manifolds. Discarding low-confidence tail samples starves the minority classes of supervision, while including noisy predictions amplifies confirmation bias and degrades representation learning.

Drawing inspiration from Schrödinger Bridge theory, which models optimal stochastic transitions between probability distributions, this work rethinks class exemplars as semantic endpoints and constructs smooth intermediate pathways in the latent space. Core idea: construct class-conditional Gaussian Feature Bridges connecting uncertain unlabeled representations to dynamic prototype anchors, regularized via bridge consistency loss and confidence-guided BridgeMix to enforce geometric smoothness on the decision boundary.

Method

Overall Architecture

The GBC framework guides uncertain representations of unlabeled instances toward high-quality class anchors maintained in a dynamic Prototype Atlas (PA). The end-to-end pipeline consists of four interconnected stages: first, maintaining a compact, class-indexed atlas of labeled exemplars and high-confidence pseudo-anchors; second, generating a class-conditional Gaussian bridge in deep latent layers and softly fusing the bridged feature into the student network via a center-gated residual update; third, establishing a geometric logarithmic target on the probability simplex to enforce bridge consistency regularization; and finally, regularizing the manifold via BridgeMix, which interpolates sample-anchor pairs weighted by pseudo-label certainty.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Unlabeled sample x_u<br/>Weak / Strong Augmentations"] --> B["Dynamic Prototype Atlas<br/>Anchor Retrieval & Maintenance"]
    B --> C["Gaussian Feature Bridge<br/>Latent Stochastic Path Interpolation"]
    C --> D["Residual Feature Fusion<br/>Center-Gated Forward Propagation"]
    D --> E["Geometric Target Alignment<br/>Simplex Log-Interpolation Loss"]
    E --> F["Confidence-Guided BridgeMix<br/>Dual Sample-Anchor Manifold Mixing"]

Key Designs

1. Dynamic Prototype Atlas Maintenance: Ensuring Clean and Diverse Class Anchors

Under severe long-tailed skew, tail classes suffer from scarce, noisy exemplars. To provide dependable target anchors, GBC dynamically maintains a class-indexed Prototype Atlas \(\mathcal{B} = \{(f_{\text{kp}}^c, c, \text{conf}, \text{src})\}\). Initialized with labeled samples, the atlas is continuously updated with teacher-predicted features whose confidence exceeds a class-adaptive threshold \(\tau_c\). To prevent head classes from dominating the memory, atlas capacity is bounded within 2%–8% of the dataset size. Redundancy is pruned via cosine-distance thresholds, and stale pseudo-anchors are evicted via temporal decay. When an extreme tail class lacks valid candidate pseudo-anchors in a given batch, the system seamlessly falls back to an exponential moving average (EMA) prototype, ensuring consistent geometric representation across both head and tail categories.

2. Latent Gaussian Feature Bridging: Controllable and Smooth Semantic Transport

To pull drifting unlabeled features back toward the true class manifold without introducing abrupt semantic shocks, GBC builds a stochastic Gaussian bridge at the student's penultimate feature layer (e.g., ResNet-50 Stage 4). Given an unlabeled feature \(f_u\) and its retrieved anchor \(f_{\text{kp}}\), the intermediate bridged state \(f_t\) is defined as:

\[f_t = (1 - g(t)) f_u + g(t) f_{\text{kp}} + \sigma(t) \varepsilon\]

where the transition parameter is sampled from \(t \sim \text{Beta}(2, 2)\) truncated to \([0.2, 0.8]\), the monotonic path function is set to \(g(t) = t\), and the stochastic diffusion noise is scaled by \(\sigma(t) = \nu \sqrt{t(1 - t)}\) with \(\varepsilon \sim \mathcal{N}(0, \mathbf{I})\). Rather than abruptly replacing the student's hidden activation, \(f_t\) is softly merged into the representation \(f_{\text{stu}}\) via a gated residual projection:

\[\tilde{f} = f_{\text{stu}} + \omega(t) \mathcal{P}(f_t - f_{\text{stu}})\]

The weighting coefficient \(\omega(t) = 4t(1 - t)\) peaks at the bridge midpoint (\(t = 0.5\)) to impose maximal regularization where semantic ambiguity is highest, while \(\mathcal{P}(\cdot)\) ensures dimensional and geometric alignment.

3. Geometric Target Construction and Bridge Consistency Loss: Regularizing Decision Boundaries

Intermediate bridged representations require supervision that respects the underlying geometry of the probability simplex rather than naive arithmetic interpolation. GBC computes an intermediate soft target \(q_t\) by geometrically interpolating the log-probabilities of the unlabeled prediction \(\hat{q}_u\) and the anchor prediction \(\hat{q}_{\text{kp}}\):

\[q_t = \text{softmax}\big((1 - t)\log \hat{q}_u + t\log \hat{q}_{\text{kp}}\big)\]

Combined with a class-frequency reweighting factor \(w_c \propto (\bar{f} / f_c)^\gamma\) that prioritizes under-represented classes, the bridge consistency loss is formulated as:

\[\mathcal{L}_{\text{bridge}} = \mathbb{E}_{x_u, t}\left[ w_c \cdot \omega(t) \cdot \text{KL}\big(q_t \parallel p_\theta(x_u^{\text{strong}}; \tilde{f})\big) \right]\]

Theoretical analysis confirms that minimizing this loss enforces local Lipschitz continuity on the student model and directly constrains the expected curvature of the decision boundary \(\mathbb{E}_t[\kappa(\partial \mathcal{M}_c)] \le C_0 + C_1 \mathcal{L}_{\text{bridge}}^{1/2}\), effectively flattening class boundaries and suppressing feature noise.

4. Confidence-Guided BridgeMix: Dual Manifold Regularization

Standard MixUp uses symmetric random weights, which often causes noisy pseudo-labels to corrupt reliable representations in imbalanced regimes. BridgeMix introduces an asymmetric interpolation scheme across sample pairs \((x_i, x_j)\) and their corresponding anchors \((f_i^{\text{anchor}}, f_j^{\text{anchor}})\), weighted by their prediction confidences \((o_i, o_j)\):

\[\lambda_i = \frac{o_i}{o_i + o_j}\]

The mixed features and anchors are formed as \(f_{u}^{\text{mix}} = \lambda_i f_i + (1 - \lambda_i) f_j\) and \(f_{\text{mix}}^{\text{anchor}} = \lambda_i f_i^{\text{anchor}} + (1 - \lambda_i) f_j^{\text{anchor}}\), which then feed into the Gaussian bridge and geometric soft target formulation. This directional guidance allows highly certain samples to stabilize uncertain ones, theoretically tightening the generalization error bound via reduced Rademacher complexity.

Loss & Training

The overall training objective combines supervised cross-entropy, unsupervised consistency, and geometric bridge consistency:

\[\mathcal{L} = \mathcal{L}_{\text{sup}} + \mu \mathcal{L}_{\text{unsup}} + \beta \mathcal{L}_{\text{bridge}}\]

Hyperparameter \(\beta\) is linearly warmed up to 0.75 in early epochs to avoid destabilizing early representation learning. Standard hyperparameters specify noise intensity \(\nu = 0.10\) and Beta distribution parameter \(\alpha = 2.0\). A small epsilon \(\epsilon = 10^{-6}\) is added during logarithmic target computation to ensure numerical stability against one-hot or peaky distributions.

Key Experimental Results

Main Results

GBC was extensively benchmarked on CIFAR10-LT under five distinct unlabeled distribution regimes (consistent, uniform, reversed, middle, head-tail) as well as large-scale ImageNet-127 and ImageNet-1K benchmarks.

Top-1 classification accuracy on CIFAR10-LT (\(N_1 = 500, M_1 = 4000\)) across unlabeled distribution regimes:

Method	consistent (\(\gamma_\ell=100, \gamma_u=100\))	uniform (\(\gamma_\ell=100, \gamma_u=1\))	reversed (\(\gamma_\ell=100, \gamma_u=1/100\))	middle (\(\gamma_\ell=100, \gamma_u=100\))	head-tail (\(\gamma_\ell=100, \gamma_u=100\))
FixMatch	65.48 ± 0.81	72.49 ± 0.61	62.53 ± 0.95	65.95 ± 0.74	60.52 ± 0.93
w/ CReST+	70.45 ± 0.63	84.96 ± 0.43	69.03 ± 0.71	72.47 ± 0.54	65.95 ± 0.76
w/ DASO	74.48 ± 0.51	89.52 ± 0.29	75.51 ± 0.52	78.53 ± 0.41	72.48 ± 0.62
w/ ACR	81.61 ± 0.33	92.09 ± 0.19	85.03 ± 0.28	73.62 ± 0.46	79.82 ± 0.39
w/ DyTrim	75.45 ± 0.36	90.12 ± 0.22	76.51 ± 0.41	78.94 ± 0.35	72.96 ± 0.42
w/ SimPro	85.66 ± 0.28	93.76 ± 0.10	85.84 ± 0.21	84.83 ± 0.20	82.98 ± 0.30
w/ SimPro + Manifold MixUp	89.14 ± 0.21	94.19 ± 0.09	86.11 ± 0.19	84.92 ± 0.21	83.63 ± 0.26
w/ GBC (Ours)	92.30 ± 0.17	94.51 ± 0.07	86.78 ± 0.23	85.49 ± 0.19	84.50 ± 0.27

Scaling evaluation on ImageNet-127 and ImageNet-1K (Table 4 of the paper):

Dataset & Resolution	FixMatch	SimPro	Meta-Expert	DyTrim	GBC (Ours)	Gain over SimPro
ImageNet-127 (\(32\times32, \gamma_t \approx 286\))	29.7 ± 0.37	59.1 ± 0.23	60.3 ± 0.22	60.8 ± 0.25	61.9 ± 0.36	+2.8%
ImageNet-127 (\(64\times64, \gamma_t \approx 286\))	42.3 ± 0.41	67.0 ± 0.22	67.4 ± 0.23	67.9 ± 0.23	68.6 ± 0.24	+1.6%
ImageNet-127 (\(64\times64, \gamma_t = 1\))	46.7 ± 0.44	63.8 ± 0.23	64.8 ± 0.23	65.7 ± 0.28	67.0 ± 0.37	+3.2%
ImageNet-1K (\(64\times64, \gamma_t = 1\))	–	25.0 ± 0.24	25.4 ± 0.24	26.1 ± 0.25	27.5 ± 0.26	+2.5%

Ablation Study

Ablation on CIFAR10-LT (\(\gamma_\ell = 100, \gamma_u = 100\)) in Table 5 illustrates the indispensable role of each design element:

Config	Top-1 Accuracy (%)	Note
Full GBC	92.3 ± 0.17	Full Gaussian bridge consistency and BridgeMix
w/o \(\mathcal{L}_{\text{bridge}}\)	86.1 ± 0.24	Removing bridge loss causes a severe drop of -6.2%
w/o noise (\(\sigma(t) = 0\))	90.2 ± 0.21	Deterministic linear bridge loses local stochastic exploration
Hard bridge (\(\tilde{f} = f_t\))	88.4 ± 0.19	Direct feature substitution disrupts latent forward representations
w/o reweighting (uniform \(w_c\))	85.3 ± 0.22	Unweighted loss fails to protect minority tail categories

Furthermore, bridge layer selection (Table 1) shows that inserting the bridge at Stage 4 of ResNet-50 yields 92.30%, compared to 87.8% at Stage 2, validating that high-level semantics are essential for bridging. In terms of efficiency (Table 8), GBC adds only +1.7% training epoch wall-clock overhead (42.8s vs 42.1s) and 3% peak memory, demonstrating strong practical scalability.

Key Findings

• Bridge consistency loss is the primary performance driver: Removing \(\mathcal{L}_{\text{bridge}}\) induces the most dramatic degradation (92.3% \(\to\) 86.1%), proving that passive anchor storage without active path-level probability alignment cannot prevent semantic drift.
• Stochastic Gaussian perturbation prevents overfitting: Adding Gaussian diffusion noise improves accuracy by +2.1% over deterministic interpolation, confirming that stochastic neighborhood exploration acts as an effective geometric regularizer.
• Superior recovery under reversed distribution shifts: Under the reversed unlabeled setting where pseudo-label noise is most extreme, GBC exhibits rapid, monotonic decay in teacher-student KL divergence, achieving the lowest final divergence after 300 epochs and proving its resilience to confirmation bias.

Highlights & Insights

• Schrödinger Bridge-inspired representation flow: Replaces heuristic logit post-processing with a grounded stochastic transport process in latent space, bridging optimal transport principles and semi-supervised consistency training.
• Asymmetric confidence-guided BridgeMix: Overcomes the vulnerability of standard MixUp in imbalanced settings by dynamically re-weighting sample and anchor interpolations according to prediction certainty.
• Rigorous geometric and generalization theory: Establishes formal mathematical proofs for local Lipschitz continuity, decision boundary flattening, and tightened generalization risk bounds via Rademacher complexity.

Limitations & Future Work

• Threshold sensitivity across diverse dataset granularities: While strict confidence thresholds \(\tau_c\) consistently enhance performance on small-label datasets like CIFAR10, they can severely curtail anchor recall on fine-grained benchmarks like CIFAR100, necessitating adaptive per-class tuning.
• Cold-start dependency on feature backbone: The quality of initial anchor prototypes relies on the initial separability of the feature extractor. Severe representation collapse in early cold-start regimes could impair the anchoring effect.
• Future Directions: Extending continuous Gaussian feature bridges to neural differential equations (Neural ODEs) or continuous consistency vector fields in multimodal imbalanced learning.

Related Work & Insights

• vs SimPro: SimPro constructs an elegant probabilistic framework for estimating pseudo-label distributions but does not enforce continuity in feature space; GBC adds class-conditional feature bridges and geometric regularization, outperforming SimPro by 1.0%–6.6% across CIFAR10-LT splits.
• vs Manifold MixUp: Manifold MixUp performs unconstrained interpolation between arbitrary latent vectors, risking off-manifold noise; GBC explicitly bridges samples to reliable class anchors and weights trajectories by confidence.
• vs DyTrim: DyTrim relies on output-level logit trimming to mitigate tail neglect; GBC demonstrates that internal geometric transport provides fundamentally stronger feature alignment and boundary stabilization.

Rating

• Novelty: ⭐⭐⭐⭐☆ (Innovative integration of Schrödinger Bridge-inspired stochastic transport into long-tailed SSL)
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Rigorous validation across five unlabeled distribution shifts, multiple resolutions, and deep ablations)
• Writing Quality: ⭐⭐⭐⭐⭐ (Well-structured presentation, clear motivation, and complete mathematical derivations)
• Value: ⭐⭐⭐⭐☆ (Offers an effective, scalable geometric regularization paradigm for realistic label-scarce learning)

Related Papers

• [ECCV 2026] Geometric Gradient Rectification for Safe Open-Set Semi-Supervised Learning
• [AAAI 2026] Sampling Control for Imbalanced Calibration in Semi-Supervised Learning
• [ECCV 2026] High-Throughput Event-Based Feature Detection and Tracking on an Embedded CPU
• [CVPR 2026] Confusion-Aware Spectral Regularizer for Long-Tailed Recognition
• [AAAI 2026] Semi-Supervised High Dynamic Range Image Reconstructing via Bi-Level Uncertain Area Masking