Gaussian-Neighbor Soft Labels

• Gaussian-neighbor soft labels are soft target distributions computed using Gaussian-weighted local neighborhood aggregation to capture uncertainty and geometry.
• They regularize learning by replacing hard labels with neighborhood-based averages, improving segmentation, noisy-label correction, and classification robustness.
• Different construction mechanisms—from spatial smoothing to graph diffusion—offer adaptable loss functions and methodologies for various machine learning tasks.

Gaussian-neighbor soft labels are soft target distributions formed by aggregating labels, predictions, or class indicators over a local neighborhood with weights that decay with distance, typically through a Gaussian kernel or through graph-diffusion operators built from Gaussian affinities. In current usage, the construction appears in several technically distinct forms: spatially varying label smoothing for segmentation, feature-space or k-NN label averaging for classification, posterior label smoothing on graphs, and probabilistic label spreading on Gaussian $k$-NN graphs (Murugesan et al., 2023, Bahri et al., 2021, Heo et al., 2024, Klees et al., 4 Feb 2026). This suggests a family of methods rather than a single standardized algorithm: the shared principle is that supervision is replaced or regularized by a local label distribution that reflects neighborhood geometry, ambiguity, and annotation uncertainty.

1. Formal definition and mathematical structure

A canonical Gaussian-neighbor construction begins with a weighted neighborhood graph. For points $x_i$ and $x_j$, a standard Gaussian affinity is

$$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$

often restricted to a $k$-nearest-neighbor graph and then symmetrized or normalized (Klees et al., 4 Feb 2026). In spatial label smoothing for segmentation, the corresponding soft label at a central pixel $p$ and class $k$ is written as

$$\tilde{y}_p^k = \frac{1}{\left|\sum_{i=1}^d w_i\right|} \sum_{i=1}^d y_i^k\, w_i,$$

so that $\tilde{y}_p^k$ is the Gaussian-weighted local class proportion in the surrounding patch (Murugesan et al., 2023).

The same idea can be expressed as neighborhood averaging in a learned representation space. In locally adaptive label smoothing, the $k$-NN label is

$x_i$0

and the final target is

$x_i$1

which mixes the original one-hot label with a local neighbor label distribution and a global uniform prior (Bahri et al., 2021).

A graph-diffusion formulation replaces direct averaging by repeated propagation. Classical label spreading updates

$x_i$2

with fixed point

$x_i$3

where $x_i$4 is derived from the Gaussian-neighbor graph (Klees et al., 4 Feb 2026). In this view, Gaussian-neighbor soft labels are diffusion-smoothed label distributions defined on the data manifold rather than purely pointwise kernel averages.

2. Construction mechanisms across learning settings

In noisy-label learning, one construction begins from the observation that during an early learning phase, learned representations of images from the same category still congregate together regardless of their noisy labels. A method based on this observation creates synthetic samples by aggregating original samples with their top-$x_i$5 nearest neighbours, with weights calculated using a mixture model learned from the per-sample loss distribution; under extreme label noise, it estimates soft targets by gradually correcting the noisy labels (Lu et al., 2024). A plausible implication is that Gaussian-neighbor soft labels arise naturally when the top-$x_i$6 aggregation is implemented with distance-decaying weights in the learned representation space.

A second construction uses the model’s own predictions rather than observed labels. Neighbor Consistency Regularization defines, for each training example, a similarity-weighted average of the temperature-scaled predictions of its nearest neighbors in feature space and penalizes the KL divergence between the prediction of the example and this neighborhood target. The loss is

$x_i$7

and is combined with the supervised loss as

$x_i$8

The published formulation uses cosine similarity, but the same architecture admits Gaussian weights in feature space without altering the underlying soft-target mechanism (Iscen et al., 2022).

A third construction is probabilistic rather than metric. Posterior label smoothing for node classification defines a node-specific soft label through the posterior

$x_i$9

with the neighborhood likelihood factorized by edgewise class-conditionals estimated from global graph statistics. The final soft label interpolates this posterior with the original one-hot label and additional uniform noise (Heo et al., 2024). Here the neighborhood need not be Gaussian, but the construction is directly compatible with Gaussian distance weighting over graph neighborhoods.

3. Loss functions for Gaussian-neighbor soft targets

The default objective for soft labels is soft cross-entropy, equivalently KL matching up to a target entropy constant. When labels are distributions rather than one-hot vectors, this objective trains the model to reproduce the full target distribution, including its uncertainty structure (Grossmann et al., 2022). That behavior is desirable when the soft label is treated as a ground-truth distribution, but it is not the only plausible semantics.

Collision cross-entropy proposes a different objective for soft class labels: $x_j$0 For network prediction $x_j$1 and target $x_j$2, the per-example loss is $x_j$3. It agrees with Shannon cross-entropy for one-hot labels, is symmetric in $x_j$4 and $x_j$5, and assigns zero training contribution to examples whose target $x_j$6 is uniform because the loss then becomes constant (Zhang et al., 2023). For Gaussian-neighbor soft labels, this distinction is consequential: with a narrow Gaussian, collision cross-entropy behaves much like ordinary cross-entropy, whereas with a broad or nearly uniform Gaussian it downweights those examples instead of forcing the model to copy their uncertainty.

In segmentation, IoU-oriented losses require a different extension to soft labels. Jaccard Metric Losses define

$x_j$7

and

$x_j$8

Both are metrics on $x_j$9, and on hard labels they are identical to the usual soft Jaccard loss. This makes them directly compatible with Gaussian-neighbor soft labels in semantic segmentation, where ordinary $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$0-style soft Jaccard is not well behaved for truly soft targets (Wang et al., 2023).

4. Representation geometry, ambiguity, and noisy supervision

The empirical motivation for Gaussian-neighbor soft labels is closely tied to representation geometry. In noisy-label classification, the early learning phase reveals that deep networks fit clean samples before memorizing mislabeled samples, and that representations of images from the same category still form local groups even when labels are corrupted. The synthetic-sample framework based on nearest neighbours reports that estimated soft targets provide a more accurate approximation to ground truth labels and that the learned representations have more separated and clearly bounded clusters; experiments are reported on CIFAR-10, CIFAR-100, Clothing1M, and Webvision (Lu et al., 2024).

Controlled experiments on label distributions show a related effect under ambiguity rather than explicit noise. In a synthetic dataset and the Mice Bone dataset, training with soft labels instead of hard labels improved Macro Accuracy and KL divergence simultaneously: on the synthetic dataset, hard-label training obtained $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$1 ACC and $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$2 KL, whereas soft-label training obtained $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$3 ACC and $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$4 KL; on Mice Bone, the corresponding values were $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$5 and $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$6 for hard labels, versus $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$7 and $$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$8 for soft labels (Grossmann et al., 2022). The same study reports that soft labels yield a more regular structure of the internal feature space.

Graph-structured node classification exhibits an analogous regularization effect. Posterior label smoothing improves the classification accuracy in 10 node classification datasets in most cases, outperforms the hard-label baseline in 67 out of 70 backbone–dataset combinations, and is reported to mitigate overfitting during training, leading to better generalization performance (Heo et al., 2024). Taken together, these findings suggest that Gaussian-neighbor soft labels do not merely smooth targets; they impose a local geometric prior on the latent representation, discouraging memorization of isolated errors and preserving ambiguity where the neighborhood structure supports it.

5. Major instantiations in current research

The same construction recurs across several domains, but the meaning of “neighbor” depends on the underlying geometry.

Setting	Neighbor structure	Soft label form
Image classification with noisy labels	top-$$w_{ij}=\exp\!\Big(-\frac{\|x_i-x_j\|^2}{2\sigma^2}\Big),$$9 nearest neighbours in learned representation space	synthetic samples and gradually corrected soft targets (Lu et al., 2024)
Feature-space consistency regularization	nearest neighbours in mini-batch feature space	similarity-weighted average of neighbors’ predictions (Iscen et al., 2022)
Semantic segmentation	local spatial patch with a discrete spatial Gaussian kernel	Gaussian-weighted local class proportion $k$0 (Murugesan et al., 2023)
Node classification	graph neighborhoods with neighborhood label distributions and posterior computation	posterior soft labels interpolated with one-hot labels (Heo et al., 2024)
Probabilistic label estimation on graphs	Gaussian $k$1-NN graph with diffusion kernel $k$2	propagated soft label estimates with uncertainty (Klees et al., 4 Feb 2026)
Predictive churn reduction	deep $k$3-NN in logit space	local $k$4-NN label mixed with a uniform prior (Bahri et al., 2021)

In segmentation, Spatially Varying Label Smoothing defines Gaussian-neighbor soft labels by convolving the one-hot mask with a discrete spatial Gaussian kernel. A constrained-optimization analysis shows that this is equivalent, up to constants, to cross-entropy plus a KL term that pulls the prediction toward the Gaussian-weighted local class proportions $k$5. The same analysis argues that SVLS lacks an explicit mechanism to balance the contribution of this constraint with the primary objective, and proposes an alternative based on equality constraints on the logits with an explicit penalty weight $k$6 (Murugesan et al., 2023).

In graph-based probabilistic labeling, Probabilistic Label Spreading constructs a sparse Gaussian $k$7-NN graph, normalizes it into a diffusion operator $k$8, and reuses each single annotation by propagating it with the heat-kernel-like solve $k$9. The resulting estimator is a Gaussian-neighbor soft-label method in the strict graph-theoretic sense, and the paper proves consistency of the estimated probability distributions even when the number of annotations per data point converges to zero (Klees et al., 4 Feb 2026).

6. Misconceptions, limitations, and evolving directions

A recurrent misconception is to equate Gaussian-neighbor soft labels with uniform label smoothing. The literature distinguishes these sharply. Uniform label smoothing assigns the same off-target mass to every class; neighbor-based schemes instead use local class distributions, graph posteriors, or Gaussian-weighted spatial proportions that vary from sample to sample and often from training stage to training stage (Bahri et al., 2021, Heo et al., 2024, Murugesan et al., 2023). This distinction matters because the induced uncertainty is structured rather than global.

The principal design variables are the neighborhood definition, the kernel bandwidth, and the reliability of the feature space in which neighbors are computed. In locally adaptive $p$0-NN smoothing, performance is reported to be relatively insensitive to $p$1 over a wide range, but the trade-off parameter $p$2 between local and global smoothing has a clear effect; in feature-space regularization for noisy labels, dependence on feature quality and batch-local neighbor graphs remains a limitation; in segmentation, larger spatial patches can oversmooth and the penalty weight $p$3 becomes critical for balancing calibration and discrimination (Bahri et al., 2021, Iscen et al., 2022, Murugesan et al., 2023).

Another limitation concerns the semantics of the loss. Standard cross-entropy assumes that the soft label itself is the distribution to be reproduced, whereas collision cross-entropy treats the soft label as a set of likely outcomes whose overlap with the prediction should be maximized. Jaccard-style objectives are yet another case, motivated by mIoU rather than by probabilistic matching (Zhang et al., 2023, Wang et al., 2023). There is therefore no single universally optimal learning rule for Gaussian-neighbor soft labels; the appropriate loss depends on whether the target encodes ambiguity, noise correction, pseudo-label confidence, or a metric-specific surrogate.

A plausible research direction is to combine Gaussian-neighbor constructions with dynamic or meta-learned label trajectories. Meta-learning methods that treat labels as learnable parameters show that dynamic soft labels can outperform fixed converged labels, can capture semantic relationships between classes, and can correct noisy annotations during optimization (Vyas et al., 2020). More general studies of soft-label learning further indicate that wrapper methods based on duplication, weighting, and ensembles often outperform hard-label methods, especially for limited sample sizes, imbalanced data, and probability estimation (Vries et al., 2024). This suggests that future Gaussian-neighbor soft labels may be less static kernels than adaptive distributions whose neighborhood geometry, uncertainty, and training objective are learned jointly.