Local-Global Consistency Loss (LGCL)

• Local-Global Consistency Loss (LGCL) is a regularization paradigm that aligns local predictions with global structures using varied formulations such as graph-based penalties, mixed-view consistency, and cycle-consistency.
• It enhances robustness and generalization across domains like EEG emotion recognition, long-tailed visual recognition, and semi-supervised sound event detection by reducing discrepancies between local signals and global context.
• Empirical studies demonstrate that LGCL improves performance metrics and stabilizes training, though challenges remain in its reliance on hand-crafted graphs and inconsistent cross-domain terminology.

Searching arXiv for papers directly relevant to “Local-Global Consistency Loss” and closely related formulations. search_arxiv({"query":"\"Local-Global Consistency Loss\" arXiv", "max_results": 10}) search_arxiv(query="\"Local-Global Consistency Loss\" arXiv", max_results=10) search_arxiv(query="(Du et al., 2023)", max_results=5) Local-Global Consistency Loss (LGCL) denotes a class of objectives that couple local predictions, local correspondences, or local region descriptors with a global summary, global path, or global representation. The surveyed arXiv literature suggests that LGCL is not a single canonical loss but a recurring design principle whose exact mathematical form depends on domain. In EEG emotion recognition, the name Local-Global Consistency Loss is explicit and refers to a graph-weighted penalty that keeps window-level predictions close to the trial-wise expected prediction (Zeng et al., 15 Jul 2025). Closely related formulations appear under other names in long-tailed visual recognition, semi-supervised sound event detection, temporal alignment, mammography, deep odometry, cross-view geo-localization, and imbalanced learning (Du et al., 2023, Li et al., 2023, Hadji et al., 2021, Chen et al., 2022, Damirchi et al., 2021, Wang et al., 25 Sep 2025, Li et al., 2022).

1. Terminology and scope

The literature uses multiple names for closely related local-global consistency mechanisms. In some papers the mechanism is an explicit loss; in others it is a pair of regularizers, a cycle-consistency term, or an architectural constraint.

Context	Official term	Form
EEG emotion recognition	Local-Global Consistency Loss	Explicit graph-based regularizer
Long-tailed visual recognition	global and local mixture consistency loss	SimSiam-style auxiliary loss
Semi-supervised sound event detection	Local and Global Consistency regularization	Local loss plus global prototype loss
Temporal alignment	global cycle consistency loss	Round-trip correspondence verification
Mammogram classification	global consistency learning	Cross-view similarity loss
Cross-view geo-localization	global-local consistency	Super-node graph mechanism
Imbalanced learning	local-global structure consistency mechanism	Local/global latent-space alignment

The paper on Global and Local Mixture Consistency cumulative learning states that its “LGCL” idea is presented under the name Global and Local Mixture Consistency Learning and, in the abstract and conclusion, as the global and local mixture consistency loss; it does not separately define a different loss called LGCL (Du et al., 2023). The paper on Semi-supervised Sound Event Detection with Local and Global Consistency Regularization likewise does not use the exact term LGCL, but instead denotes a local consistency loss $\mathcal{L}_{\text{CLC}}$ and a global consistency loss $\mathcal{L}_{\text{PGC}}$ inside an overall LGC framework (Li et al., 2023). The cross-view geo-localization paper is even more explicit that its mechanism is conceptually comparable to LGCL but is not named LGCL and is not formulated as a standalone explicit consistency loss; its consistency is primarily built into a graph/message-passing architecture with a virtual super node (Wang et al., 25 Sep 2025).

A common misconception is therefore to treat LGCL as a single standardized formula. The surveyed papers instead indicate a family resemblance: a local entity is first defined, a global entity is then formed or inferred, and consistency is enforced either by optimization or by architecture.

2. Recurrent mathematical structure

The surveyed work suggests a recurring decomposition into local object, global object, and coupling mechanism. The local object may be a 1-second EEG window, a mixed image view, a spectrogram frame embedding, a frame-to-frame pose, a local mammographic region, or a patch node in a graph. The global object may be a trial-level expected prediction, a second mixed view of the same mini-batch, a class prototype set, a monotone alignment path, an integrated trajectory, or a whole-image scene summary.

The coupling mechanism varies substantially. In the EEG formulation, each local prediction is penalized against the trial-wise average using the pseudoinverse of a graph Laplacian: $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$ In GLMC for long-tailed recognition, the coupling is a SimSiam-style negative cosine similarity between a global MixUp branch and a local CutMix branch: $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$ In temporal alignment, local frame-to-frame matching probabilities are embedded in a DTW-style path and then checked by a global cycle-consistency condition,

$$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$

with loss

$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$

These formulations do not share a single algebraic template, but they repeatedly instantiate the same design idea: local behavior is not treated as independently supervised, and global structure is not left implicit (Zeng et al., 15 Jul 2025, Du et al., 2023, Hadji et al., 2021).

This suggests that LGCL is best understood as a structural regularization principle. Depending on the task, “global” may mean temporal expectation, cross-view agreement, integrated pose, or scene-level semantics, while “local” may refer to segments, patches, regions, or incremental motions.

3. Explicit LGCL in EEG emotion recognition

The clearest named LGCL appears in Commuting Distance Regularization for Timescale-Dependent Label Inconsistency in EEG Emotion Recognition (Zeng et al., 15 Jul 2025). The motivation is Timescale Dependent Label Inconsistency (TsDLI): datasets such as DREAMER and DEAP provide one label per trial after a long stimulus, while practical models operate on short windows. Copying the same trial label to every 1-second segment creates a mismatch because local emotion may fluctuate within the trial even when the trial-level self-report remains valid.

The paper formalizes the global label through the Full Expectation Assumption (FEA),

$Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$

and interprets the global prediction as an expectation over local predictions. LGCL then penalizes the deviation of each local prediction $\hat{Y}(t_i)$ from the trial-wise average $E[\hat{Y}(t)]$ in graph space. The graph is an undirected emotion transition graph $G(V,E)$, and for dimensional emotion ratings the default graph is a line graph,

$\mathcal{L}_{\text{PGC}}$0

Its Laplacian is

$\mathcal{L}_{\text{PGC}}$1

and the relevant geometry is defined by the commute time distance

$\mathcal{L}_{\text{PGC}}$2

The resulting LGCL is

$\mathcal{L}_{\text{PGC}}$3

The paper explicitly characterizes this as a variance-like penalty in graph space. The use of $\mathcal{L}_{\text{PGC}}$4 means that deviations are measured according to emotion-transition geometry rather than plain Euclidean distance between one-hot labels.

LGCL is paired with Local Variation Loss (LVL),

$\mathcal{L}_{\text{PGC}}$5

which regularizes adjacent windows rather than local-to-global agreement. The appendix states that under certain conditions the two are equivalent up to constants,

$\mathcal{L}_{\text{PGC}}$6

and the paper also evaluates mixtures

$\mathcal{L}_{\text{PGC}}$7

No statistically significant differences are reported across the tested mixtures, which the authors interpret as supporting the equivalence result.

In training, each trial is preprocessed, split into 1-second non-overlapping windows, passed through a backbone such as EEGNet, LSTM, or a Transformer-based model, and optimized with a base task loss plus regularization. Trials are normalized using baseline z-scoring,

$\mathcal{L}_{\text{PGC}}$8

then scaled to $\mathcal{L}_{\text{PGC}}$9. The first 40 seconds are used for training and the last 20 seconds for testing. The paper injects synthetic label noise of 20% and 40%, and evaluates F1, Top-2 accuracy, $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$0, $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$1, and $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$2.

The reported empirical pattern is that LGCL consistently improves qualitative consistency, with lower $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$3, lower $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$4, and lower $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$5, while usually preserving or slightly improving quantitative performance. The paper states that LGCL often ranks second best, while LVL achieves the best aggregate rank. It also notes four limitations: LGCL regularizes predictions but does not explicitly infer true time-resolved emotion labels; the graph is hand-crafted; the study is limited to subject-dependent settings; and public datasets lack dense local ground truth.

4. Recognition losses based on mixed views and prototypes

In long-tailed visual recognition, the closest LGCL analogue is the global and local mixture consistency loss inside GLMC (Du et al., 2023). For each input batch, the method creates two mixed views from the same original data. The global mixture is produced by MixUp: $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$6 The local mixture is produced by CutMix: $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$7 with mixed labels described as “the same as MixUp.” The encoder produces representations $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$8, a projection head maps them to $$L = \frac{1}{N}\sum_{i=1}^{N} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr)^T L^{\dagger} \Bigl(\hat{Y}(t_i) - E[\hat{Y}(t)]\Bigr).$$9, and a predictor maps them to $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$0. Similarity is defined by normalized negative cosine similarity,

$\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$1

and the auxiliary loss is

$\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$2

The paper emphasizes that this is not a contrastive loss with negatives, does not require a memory bank or momentum encoder, and uses the prediction head only for this auxiliary term.

This loss is embedded in a one-stage supervised objective,

$\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$3

where $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$4 is the ordinary mixed-label cross-entropy, $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$5 is a class-frequency reweighted version, and $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$6 controls the consistency term. On CIFAR100-LT (IF=100), the paper reports that the best setting is $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$7 and improves performance by +2.41% over $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$8. In the component ablation, “with consistency, no cumulative class-balanced learning” yields 50.11 versus 38.3 for “no consistency, no cumulative class-balanced learning,” an average gain of 11.81% as reported by the authors. The full model reaches 55.88.

In semi-supervised sound event detection, Local and Global Consistency (LGC) regularization uses two separate losses rather than a single monolithic LGCL (Li et al., 2023). The local component, $\pounds_{sim}=sim(u_g,sg(h_l))+sim(u_l,sg(h_g)).$9, is a CutMix-based label-level consistency term: $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$0 The global component, $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$1, is a prototypical contrastive loss over projected frame features and multiple class prototypes: $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$2 Its full objective is

$$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$3

The reported hyperparameters are $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$4, $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$5, $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$6, $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$7, $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$8, and $$\mathbf{P}_{Y,X}\mathbf{P}_{X,Y} \approx \mathbf{I},$$9. Training uses $\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$0 for the first 100 epochs and $\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$1 for the last 100 epochs. On DESED, the paper reports 48.8 EB-F1, 0.392 PSDS$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$2, and 0.579 PSDS$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$3 for LGC versus 42.3, 0.347, and 0.545 for the baseline, and its ablation shows 44.5 EB-F1 without global consistency and 45.6 without local consistency.

These two papers exemplify two distinct uses of local-global consistency. GLMC enforces consistency between two strong mixed views of the same mini-batch to improve representation robustness under long-tailed imbalance. LGC in sound event detection explicitly separates local prediction consistency from global feature-prototype alignment.

5. Cycle-consistency, composition, and cross-view agreement

Several formulations move beyond classification regularization and define local-global consistency through alignment, composition, or cross-view agreement. In Representation Learning via Global Temporal Alignment and Cycle-Consistency, the local object is a frame-to-frame matching probability and the global object is a feasible monotone alignment path between two sequences (Hadji et al., 2021). Pairwise costs are defined as contrastive negative log probabilities,

$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$4

and accumulated through a smooth DTW recursion,

$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$5

The global cycle-consistency term checks whether mapping $\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$6 and back returns to the same element,

$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$7

with loss

$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$8

The total objective is

$\mathcal{L}_{\text{GCC}(\mathbf{X},\mathbf{Y}) = -\sum_{i=1}^M \log\big((\mathbf{P}_{Y,X}\mathbf{P}_{X,Y})_{i,i}\big).$9

with $Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$0, $Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$1, $Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$2, $Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$3, and $Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$4.

In deep odometry, the local object is the incremental relative pose and the global object is the composed pose over a short window. A Consistency-Based Loss for Deep Odometry Through Uncertainty Propagation trains on both incremental and composed motion, but weights each term by estimated uncertainty rather than fixed coefficients (Damirchi et al., 2021). The composed/global pose is obtained by SE(3) composition,

$Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$5

and its covariance is derived by propagating the incremental covariances through Lie-group composition using the Baker–Campbell–Hausdorff formula and the adjoint representation. The paper reports improvements over UA-VO of 19.8% in translation accuracy and 41.5% in rotation accuracy, and an average out-of-range rate of about 1.53% for predicted uncertainty intervals.

In multi-view mammogram classification, local-global consistency is expressed as cross-view agreement between the CC and MLO views. Multi-view Local Co-occurrence and Global Consistency Learning Improve Mammogram Classification Generalisation defines global features $Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$6, maps each view into the other’s feature space, and applies a cosine-similarity-based global loss,

$Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$7

where

$Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$8

A local co-occurrence module separately models cross-view region interactions by attention. On ADMANI-1, BRAIxMVCCL achieves 0.948 versus 0.889 for GMIC and 0.886 for EfficientNet-b0; on ADMANI-2, it reaches 0.926 versus 0.844 for GMIC and 0.838 for EfficientNet-b0 (Chen et al., 2022).

These papers illustrate that local-global consistency can be defined over temporal correspondences, integrated geometric states, or paired medical views. The mathematical device changes, but the design objective remains to prevent locally plausible predictions from becoming globally incoherent.

6. Architectural variants, empirical behavior, and limitations

Not every local-global consistency method is an explicit loss. Enhancing Cross-View Geo-Localization Generalization via Global-Local Consistency and Geometric Equivariance implements the principle through a graph with a virtual super node rather than a standalone $Y_G = \int_t p(t) Y(t)\,dt = E_t[Y(t)],$9 (Wang et al., 25 Sep 2025). Patch features become graph nodes,

$\hat{Y}(t_i)$0

a super node is initialized by global average pooling,

$\hat{Y}(t_i)$1

and message passing

$\hat{Y}(t_i)$2

allows global semantics to be aggregated and redistributed. The total training loss is the sum of InfoNCE contrastive loss and cross-entropy loss. The paper states that its evidence is module-level rather than a clean isolation of the super node alone.

A related structural variant appears in imbalanced learning. Envelope imbalanced ensemble model with deep sample learning and local-global structure consistency defines a Local-Global Structure Consistency Mechanism (LGSCM) rather than a conventional LGCL (Li et al., 2022). Its local component, the Local Manifold Structure Metric (LMSM),

$\hat{Y}(t_i)$3

preserves neighborhood structure, while the Global Structure Distribution Metric (GSDM),

$\hat{Y}(t_i)$4

aligns global distributions. These are combined with a low-rank constraint inside DSEN-LG. The paper’s ablation reports that the proposed method outperforms both Bagging+None and MlFCM on AUC, F-measure, G-mean, and MCC.

Across the surveyed applications, empirical findings repeatedly associate local-global consistency with robustness, generalization, and stability, but the supporting evidence is heterogeneous. In long-tailed recognition, the authors report that the consistency term alone raises CIFAR100-LT (IF=100) performance from 38.3 to 50.11, and the full method reaches 55.88 (Du et al., 2023). In semi-supervised SED, removing global consistency drops EB-F1 from 48.8 to 44.5 (Li et al., 2023). In mammography, the combined local and global design improves a fusion baseline from 0.9088 to 0.9478 AUC (Chen et al., 2022). In odometry, the global composed loss is effective specifically because its weight is adapted by propagated uncertainty rather than hand tuning (Damirchi et al., 2021).

The literature also identifies several limitations. Terminology is inconsistent, and in some papers the mechanism is architectural rather than loss-based. Some formulations rely on hand-crafted graphs, as in EEG emotion recognition. Some empirical evidence is only at module level, as in the geo-localization super-node model. In several settings, especially where dense local ground truth is unavailable, evaluation of local-global coherence remains indirect. A plausible implication is that “LGCL” is most precise when used paper-specifically and less precise when used as a cross-domain label.

Taken together, the arXiv literature portrays Local-Global Consistency Loss not as one fixed equation but as a broad regularization paradigm: local predictions or descriptors are constrained by a global structure that the model must also represent. Whether implemented through graph geometry, mixed-view agreement, prototypes, cycle-consistency, trajectory composition, or super-node message passing, the central objective is to reduce the gap between local supervision signals and the global organization required by the task.