Data Fusion Block Overview

• Data Fusion Block is a modular component that combines diverse measurements or embeddings into a single, downstream-consumable representation.
• It employs varied techniques—from Bayesian update mechanisms to bilinear tensor fusion and non-parametric operations—to address specific fusion tasks.
• The design integrates both statistical and neural approaches to handle missing modalities and improve localization, classification, and tracking performance.

In the literature surveyed here, a data fusion block is a modular computational stage that combines heterogeneous measurements, modality-specific embeddings, or partial posterior estimates into a single fused posterior, event hypothesis, feature map, or latent representation. The term is used across probabilistic sensor fusion, multimodal deep learning, and operational analytics, but it does not denote a single standardized operator. Instead, it refers to a family of placement-specific modules ranging from Bayesian update-and-resample schemes and block-sparse evidence maximization to bilinear tensor fusion, outer-arithmetic interaction maps, attention-based aggregation, state-space fusion, and microservice-based spatio-temporal correlation engines (Miller et al., 2020, Ben-Younes et al., 2019, Ortiz et al., 2020, Liu et al., 2022, Alwazzan et al., 2024, Peng et al., 2024, Farooq et al., 2024, Alwazzan et al., 2024, Möderl et al., 17 Mar 2025, Jing et al., 26 Jul 2025, Dastagir et al., 6 May 2026).

1. Scope, terminology, and architectural placement

The surveyed papers suggest that a data fusion block is best understood by its interface and placement. In particle-based Bayesian fusion, the block takes $K$ empirical particle approximations and likelihood functions and returns a fused empirical posterior approximation to $P(\theta\mid z_1,\dots,z_K)$. In a common-operating-picture framework, it sits between modality-specific event generators and analyst-facing services, producing geolocated fused event hypotheses from AIS and FMV streams. In multimodal neural systems, it typically consumes modality-specific latent vectors or feature maps and emits a shared representation for classification, segmentation, depth estimation, or survival prediction (Miller et al., 2020, Ortiz et al., 2020, Liu et al., 2022, Alwazzan et al., 2024, Peng et al., 2024, Farooq et al., 2024).

Family	Inputs	Fused output
Particle posterior fusion	Local particle sets and likelihoods	Empirical posterior
Event-correlation fusion	AIS_Event and FMV_Event	Fused_Event
Bilinear or outer-arithmetic fusion	Two latent vectors	Joint latent or logits
N-to-One attention fusion	Available modality feature maps	Shared feature map
Dual-input state-space fusion	Spatial and spectral feature maps	Fused feature map

Placement varies systematically with task structure. MOAB is described as an intermediate or late fusion block fed by modality-specific embeddings from a histopathology backbone and a genomics MLP, while the precision-oncology MOAB is placed at the end of a dual-fusion pipeline after early fusion and gated MIL attention. SFusion is inserted between upstream encoders and downstream decision networks. FusionMamba appears at every stage of the spectral U-Net, where it injects spatial detail into spectral features hierarchically. In BARFI-Q, the Fusion Block follows two BAR-Transformer branches and precedes the Quantum Feature Mapping module (Alwazzan et al., 2024, Alwazzan et al., 2024, Liu et al., 2022, Peng et al., 2024, Dastagir et al., 6 May 2026).

2. Bayesian and statistical formulations

A probabilistic data fusion block may act directly on posterior measures rather than learned features. In "Bayesian Fusion of Data Partitioned Particle Estimates" (Miller et al., 2020), each sensor $j$ provides an empirical approximation

$$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$

with local likelihood $g_j(\theta)=P(z_j\mid \theta)$. The norm-together fusion rule assigns pooled multiple-importance-sampling weights

$$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$

then resamples $N$ particles from the combined pool. After resampling,

$$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$

Under bounded-likelihood assumptions $g_j(\theta)\in[\kappa,\kappa^{-1}]$, the paper states

$d(\pi^N,\pi)=O(1/\sqrt{N}),$

where $P(\theta\mid z_1,\dots,z_K)$0 is the $P(\theta\mid z_1,\dots,z_K)$1 root-mean-square distance on measures.

A distinct Bayesian formulation appears in the block-sparse SBL method for multi-sensor data fusion (Möderl et al., 17 Mar 2025). There, each sensor $P(\theta\mid z_1,\dots,z_K)$2 produces a complex measurement vector $P(\theta\mid z_1,\dots,z_K)$3 governed by

$P(\theta\mid z_1,\dots,z_K)$4

Fusion is imposed through a row-sparse prior shared across sensors,

$P(\theta\mid z_1,\dots,z_K)$5

so that all $P(\theta\mid z_1,\dots,z_K)$6 share the same sparsity pattern. Inference follows a Type-II evidence-maximization objective over $P(\theta\mid z_1,\dots,z_K)$7, coordinate updates in $P(\theta\mid z_1,\dots,z_K)$8, and an EM update for the noise precision $P(\theta\mid z_1,\dots,z_K)$9. The block outputs estimated object positions $j$0, per-sensor complex amplitudes $j$1, and $j$2, explicitly positioning fusion before track-level Bayesian MOT.

These two formulations exemplify a common pattern: fusion occurs in the posterior or likelihood machinery itself rather than through a downstream heuristic combination. This suggests a sharp distinction between data fusion blocks that preserve measure-theoretic semantics and those that primarily learn a discriminative latent representation.

3. Bilinear, outer-arithmetic, and non-parametric feature fusion

A large class of neural data fusion blocks is designed around explicit cross-modal interactions. BLOCK uses a block-superdiagonal tensor decomposition to structure a third-order bilinear parameter tensor $j$3 as

$j$4

or equivalently as a sum of $j$5 block terms. Inputs $j$6 and $j$7 are projected into block space, fused blockwise, concatenated, and mapped back to $j$8. The construction interpolates between CP and Tucker extremes through the block-term rank $j$9, yielding a controllable tradeoff between expressiveness and parameter efficiency (Ben-Younes et al., 2019).

MOAB replaces bilinear cores with four outer-arithmetic interaction branches. In the brain-tumor grading formulation, $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$0 are extended to $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$1, $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$2, $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$3, and $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$4, and four $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$5 maps are computed: outer product, outer division, outer addition, and outer subtraction. A sigmoid is applied elementwise to each branch, the four maps are stacked into $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$6, and a single $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$7 2D convolution merges channels before flattening and classification (Alwazzan et al., 2024). In the precision-oncology formulation, the same logic is scaled to $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$8, generating $$\pi_{j,0}^N(d\theta)=\sum_{i=1}^N \frac{1}{N}\,\delta_{\theta_j(i)}(d\theta),$$9 outer-arithmetic maps, stacking them into $g_j(\theta)=P(z_j\mid \theta)$0, applying a $g_j(\theta)=P(z_j\mid \theta)$1 convolution, flattening, and passing the result to a classifier or survival head (Alwazzan et al., 2024). In both versions, appending $g_j(\theta)=P(z_j\mid \theta)$2 or $g_j(\theta)=P(z_j\mid \theta)$3 is used to preserve pure-modality terms along entire rows and columns.

A contrasting design is the non-parametric fusion layer in the head-and-neck cancer survival model (Farooq et al., 2024). After 3D-ResNet-50 and CBAM processing of CT and PET inputs, the fusion block takes two modality vectors $g_j(\theta)=P(z_j\mid \theta)$4, stacks them into $g_j(\theta)=P(z_j\mid \theta)$5, and computes the fused representation by row-wise maximum: $g_j(\theta)=P(z_j\mid \theta)$6 element-wise. The paper explicitly states that the fusion block itself is non-parametric and contains no weights or biases.

Taken together, these designs show that the term "data fusion block" covers both high-order parametric interaction models and parameter-free selection rules. A plausible implication is that the choice of block reflects whether the task is expected to benefit more from enumerating cross-modal structure or from imposing a conservative competition between modalities.

4. Attention, Transformer, and state-space fusion blocks

Attention-based fusion blocks treat modalities as correlated token sets or feature streams. SFusion is an N-to-One multimodal fusion block comprising Correlation Extraction through stacked Transformer-style self-attention layers and Modal Attention through a voxel-wise softmax across the available modalities (Liu et al., 2022). If the available modalities are indexed by $g_j(\theta)=P(z_j\mid \theta)$7, their feature maps are flattened and concatenated into

$g_j(\theta)=P(z_j\mid \theta)$8

processed by multi-head self-attention, then reshaped back to modality-wise maps. The final fused map is

$g_j(\theta)=P(z_j\mid \theta)$9

Because both stages operate only on modalities actually present in $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$0, the block natively supports missing modalities without zero-padding or synthesized substitutes.

FuseFormer adopts a two-branch multi-scale fusion block for visual and thermal image fusion (Erdogan et al., 2024). At each scale, the visual and infrared features are processed by a small CNN branch with two $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$1 convolutions and by an axial-attention branch with sequential height- and width-axis attention. The original modality features and both branch outputs are concatenated along channels and reduced back to the original width by a $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$2 convolution, BN, and ReLU. Its stage-2 objective combines a feature consistency term with a symmetric SSIM term,

$$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$3

where $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$4 penalizes lack of structural similarity to both input modalities rather than only the visual image.

FusionMamba extends single-input Mamba to a dual-input Fusion State Space Model (Peng et al., 2024). Two feature maps $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$5 are normalized and projected, flattened along four raster-scan directions, and passed through two FSSM streams that cross-condition the state-space parameters on the other input. After unflattening and summing over directions, each stream is gated by $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$6, added residually, and combined through a final $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$7 convolution. The details emphasize linear complexity in $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$8 together with global receptive field through full-sequence selective scans.

UniCT Depth introduces the Convolution-compensated ViT Dual SA block, which combines Context Modeling Self-Attention, Modal Fusion Self-Attention, and a Detail Compensation Convolution module (Jing et al., 26 Jul 2025). Patch-embedded tokens are processed by CMSA and MFSA in parallel, both outputs are modulated by a DCC-derived attention map, concatenated, and merged by two successive $$w_j(i)= \frac{\prod_{\ell\neq j} g_\ell(\theta_j(i))} {\sum_{j'=1}^K\sum_{i'=1}^N \prod_{\ell\neq j'} g_\ell(\theta_{j'}(i'))},$$9 convolutions with LayerNorm and GELU. The DCC branch uses channel pooling, a spatial attention map, and two $N$0 convolutions to recover edge and texture detail.

BARFI-Q places a Fusion Block between two temporal branches and a quantum feature-mapping stage (Dastagir et al., 6 May 2026). The branch outputs are concatenated, linearly aligned to a common latent dimension, refined by multiscale channel attention with parallel $N$1, $N$2, and $N$3 convolutions, refined again by multiscale spatial attention over temporal tokens, and projected to the final fused latent $N$4. The residual additions are weighted by learned channel and token attention maps rather than fixed additive paths.

5. Systems-level fusion engines and integration into larger pipelines

Not all data fusion blocks are neural layers. In the VASP common-operating-picture framework, the Data Fusion Block is a microservice architecture organized into Ingestion and Preprocessing, modality-specific Event Generation, a Cross-Modal Correlation and Fusion Engine, and Consolidation, Logging, and API services (Ortiz et al., 2020). The AIS branch includes geofence, disappearance/appearance, course-deviation, proximity, and event-store components. The FMV branch includes deep-learning detection, georeferencing, activity heuristics, and an FMV event store. The fusion engine performs spatio-temporal joining, vessel verification and tip-and-cue, fused event generation, and confidence scoring

$N$5

Outputs are logged, exposed through REST endpoints, and published into a common operating picture.

Within neural pipelines, integration strategy is itself a central design variable. MOAB is explicitly described as a late fusion point that preserves rich unimodal feature learning while introducing cross-modal interactions immediately before classification (Alwazzan et al., 2024). The precision-oncology system uses a dual-fusion design in which omic embeddings are projected onto WSI patches in latent space for early fusion, refined by a Multiple Instance Learning gated attention mechanism, and then reintroduced at slide level through MOAB for late fusion (Alwazzan et al., 2024). SFusion is presented as a plug-in replacement for existing fusion modules in both human activity recognition and brain-tumor segmentation backbones (Liu et al., 2022). The CT/PET survival model feeds the fused vector directly into a fully parametric discrete-time survival head, while FusionMamba injects spatial detail into the spectral U-Net at all five scales rather than only once at the end (Farooq et al., 2024, Peng et al., 2024).

These examples indicate that "fusion block" may describe either a local operator inside a deep stage or a pipeline-level correlator spanning multiple services. The surveyed literature therefore resists a narrow definition tied only to tensor fusion.

6. Computational properties, empirical behavior, and recurrent misconceptions

Computational and numerical behavior differ sharply across formulations. The particle-fusion block has cost $N$6, requires a pool of $N$7 particles, and is accompanied by explicit recommendations for log-domain computation, subtraction of the maximum log-weight before exponentiation, flooring $N$8 to $N$9, and ESS-triggered resampling with Gaussian perturbation to mitigate degeneracy (Miller et al., 2020). The paper also notes that multinomial resampling is simplest, whereas systematic or stratified resampling gives lower variance at the same $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$0 cost. In the brain-tumor MOAB, each outer branch requires $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$1 scalar operations, the $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$2 convolution adds negligible cost, the dominant cost remains the backbone, and stable convergence is attributed to layer normalization, mild dropout, orthogonal initialization, and a learning rate of $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$3 for fusion with weight decay $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$4 (Alwazzan et al., 2024). FusionMamba explicitly contrasts its linear $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$5 complexity with the quadratic $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$6 term of self- or cross-attention over image tokens (Peng et al., 2024). By contrast, the CT/PET max fusion layer is stated to introduce no additional learned parameters inside the fusion block itself (Farooq et al., 2024).

Reported outcomes are likewise heterogeneous. In the multi-sensor Keplerian orbit determination example, norm-together particle fusion yielded final fused position RMSE $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$7 km, compared with $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$8 km and $$\pi^N(d\theta)=\frac{1}{N}\sum_{n=1}^N \delta_{\theta(n)}(d\theta), \qquad \pi^N(f)=\frac{1}{N}\sum_{n=1}^N f(\theta(n))\approx E[f(\theta)\mid z_{1:K}].$$9 km for the individual sensors; in bearings-only sequential tracking, fused trajectories tracked ground truth with $g_j(\theta)\in[\kappa,\kappa^{-1}]$0 rad error while single-sensor posteriors often degenerated (Miller et al., 2020). In glioma grading, MOAB with ConvNeXt+MLP achieved $g_j(\theta)\in[\kappa,\kappa^{-1}]$1-micro $g_j(\theta)\in[\kappa,\kappa^{-1}]$2 and $g_j(\theta)\in[\kappa,\kappa^{-1}]$3-macro $g_j(\theta)\in[\kappa,\kappa^{-1}]$4, exceeding simple concatenation and the cited prior fusion baselines; the paper also reports superior t-SNE cluster separation for Grades II, III, and IV (Alwazzan et al., 2024). SFusion improved SHL2019 overall accuracy from $g_j(\theta)\in[\kappa,\kappa^{-1}]$5 to $g_j(\theta)\in[\kappa,\kappa^{-1}]$6 and improved BraTS2020 scores across WT, TC, and ET under 15 real-missing combinations (Liu et al., 2022). In the multi-radar SBL setting, fusion from $g_j(\theta)\in[\kappa,\kappa^{-1}]$7 radars reduced the false-alarm rate by roughly an order of magnitude at fixed detection threshold and improved localization RMSE from $g_j(\theta)\in[\kappa,\kappa^{-1}]$8 m for $g_j(\theta)\in[\kappa,\kappa^{-1}]$9 to $d(\pi^N,\pi)=O(1/\sqrt{N}),$0 m for $d(\pi^N,\pi)=O(1/\sqrt{N}),$1 (Möderl et al., 17 Mar 2025).

The literature surveyed here counters several common simplifications. First, a data fusion block is not necessarily a learned parametric layer: the CT/PET max operator is non-parametric, and the VASP fusion engine relies on statistical prediction, heuristic activity logic, and spatio-temporal nearest-neighbor search rather than end-to-end representation learning (Farooq et al., 2024, Ortiz et al., 2020). Second, a fusion block does not necessarily assume a fixed modality count: SFusion is explicitly formulated for the N-to-One setting with missing modalities (Liu et al., 2022). Third, the term is not restricted to images or to classification heads; it includes posterior fusion, event correlation, trajectory fusion, localization, and common-support Bayesian recovery (Miller et al., 2020, Möderl et al., 17 Mar 2025). This suggests that the most stable encyclopedia-level definition is functional rather than architectural: a data fusion block is the module that converts multiple partial evidential sources into a single downstream-consumable representation or hypothesis.