GRC-Net: Structured Neural Models

• GRC-Net is a framework integrating multi-modal inputs with global context via specialized fusion and attention modules to boost performance.
• Its variants employ dual-branch architectures, graph revision, and global receptive convolutions to address challenges in segmentation, recommendation, and node classification.
• Empirical results demonstrate significant improvements in metrics (up to 18 mIoU points and 10–18% gains) while keeping computational overhead low.

GRC-Net

GRC-Net is a designation used for several distinct neural network architectures, each situated in a different research domain but sharing an emphasis on introducing additional structure, global context, or explicit revision/fusion mechanisms into standard convolutional or graph-based models. This entry summarizes four prominent instantiations of GRC-Net, encompassing LiDAR segmentation under adverse conditions (Geometry-Reflectance Collaboration Network), graph-revised and graph-refined convolutional frameworks, semantic segmentation networks with global receptive convolution, and a Gram Residual Co-attention Net for EEG-based epilepsy prediction. Despite architectural and task-specific differences, these networks are unified by their focus on enhancing robustness and discriminative power through explicit, often multi-level, feature interaction, structured attention or revision layers, and hybrid local-global modeling.

1. Geometry-Reflectance Collaboration Network for LiDAR Segmentation

The Geometry-Reflectance Collaboration Network (GRC-Net) targets the challenge of semantic segmentation of LiDAR point clouds under adverse weather, explicitly modeling the heterogeneous domain shifts between geometric and reflectance modalities. GRC-Net employs a dual-branch structure with separate sparse 3D convolution (for geometry) and 2D convolutional (for reflectance) encoders. Early disentanglement is followed by a multi-level feature collaboration module that fuses complementary cues and suppresses weather-induced noise (Yang et al., 3 Jun 2025).

For each point in the LiDAR scan $P=\{(x,y,z,r)\}$, the geometry branch voxelizes $(x,y,z)$ and processes it through a MinkNet-style sparse 3D ConvNet, while the reflectance branch projects $r$ to a 2D range image processed via a MobileNetV2-style depthwise separable ConvNet. Both feature streams are passed through a stochastic information bottleneck characterized by learned per-point Gaussian distributions. The collaboration module optimizes a complementarity-aware information constraint (CIC) loss, which regularizes each modality's feature distribution toward a standard normal while encouraging mutual information flow between streams and suppressing redundancy.

Multi-level fusion comprises:

• Local fusion, where features are reliability-weighted based on their predicted uncertainty (channel-averaged standard deviation of the Gaussian representation).
• Global fusion, which involves cross-attention: global-query tokens aggregate non-local context from the reflectance map and inject this distilled information into the geometry branch.

A shallow sparse 3D decoder predicts semantic labels from the concatenated local/global fused features.

Experiments on SemanticKITTI→SemanticSTF, SynLiDAR→SemanticSTF, and SemanticKITTI→SemanticKITTI-C demonstrate that GRC-Net attains higher mIoU in all adverse conditions tested, outperforming RDA and other state-of-the-art baselines by up to 18 points, and providing substantial robustness across varied weather phenomena at a modest computational cost increase (Yang et al., 3 Jun 2025).

2. Graph-Refined and Graph-Revised Convolutional Networks

a) Graph-Refined Convolutional Network for Multimedia Recommendation

GRC-Net in recommendation refers to the Graph-Refined Convolutional Network, designed to enhance the robustness of GCNs in implicit-feedback recommendation scenarios by adaptively refining bipartite user-item interaction graphs (Yinwei et al., 2021). Implicit feedback graphs inevitably contain false positives—noisy, non-preference interactions—which pollute neighborhood aggregation.

The refinement layer projects multimodal item features into a metric space and iteratively routes user prototypes using neighbor attention. For each user, modality-specific preference prototypes are generated by soft-attentively aggregating neighborhood features. Edge confidence scores are then estimated bi-directionally (user←item and item←user) for each modality and fused via weighted max pooling, using trainable modality preference vectors. Low-confidence (noisy) edges are soft-pruned. Weighted GCN propagation then produces collaborative embeddings, which are concatenated with the learned user/item prototypes. The recommendation score is defined as the inner product of the final user and item representations.

GRC-Net significantly improves top-K recommendation metrics on Movielens, TikTok, and Kwai datasets, offering 10–18% gains in Recall@10 and NDCG@10 over strong content/GNN baselines. Ablation reveals the superiority of soft-pruning with prototype-driven fusion and the necessity of multi-step routing for prototype refinement (Yinwei et al., 2021).

b) Graph-Revised Convolutional Network for Node Classification

The Graph-Revised Convolutional Network (GRCN) (Yu et al., 2019) introduces a learnable graph revision module to conventional GCNs for cases where input graphs are incomplete or noisy. As opposed to solely reweighting or fully parameterizing the adjacency, GRCN produces a revised adjacency $\widetilde{A}=A+S^{(K)}$ by first embedding nodes via a GCN, then computing a dense similarity matrix (typically using a dot-product kernel), sparsifying it via top-K row selection and symmetrization, and finally adding it to the original adjacency. This revised adjacency is employed by a downstream GCN classifier.

GRCN’s theoretical framework connects to multigraph belief propagation, allowing label or feature propagation over both observed and inferred (feature-based) connections. Empirically, GRCN and its scalable variant (Fast-GRCN) demonstrate systematic gains in node classification under both edge incompleteness and label scarcity across transductive benchmarks (e.g., Cora, Citeseer, PubMed), surpassing GCN, GAT, and other reweighting-based models, with the most pronounced improvements evident when the observed graph is extremely sparse (Yu et al., 2019).

3. Global Receptive Convolution Networks (Semantic Segmentation)

Within dense prediction, GRC-Net refers to an FCN variant (FCN+), where the core innovation is the Global Receptive Convolution (GRC) (Ren et al., 2023). Standard local convolutions limit context aggregation. GRC increases the effective receptive field without extra parameters or computation by splitting channels at each layer into local and global groups:

• Local channels receive conventional $K\times K$ neighborhood convolution.
• Global channels are grouped and their convolutional grids are shifted according to channel index, enabling each global group to aggregate features across nonlocal, spatially separated grid offsets.

This mechanism grants even small-kernel convolutions direct access to global scene context, markedly improving segmentation accuracy. In FCN+, GRC blocks replace the $3\times 3$ convolutions in Stage 5 of a ResNet backbone, preserving inference speed and parameter count.

On Cityscapes, PASCAL VOC 2012, and ADE20K, FCN+ outperforms its FCN (ResNet-101) baseline and rivalling more complex context modules (e.g., Non-local, DeepLabV3+), reaching $45.74\%$ mIoU on ADE20K with no additional computational cost (Ren et al., 2023). Ablation confirms optimality of the half-local/half-global split and the insertion at late-stage convolutions.

4. GRC-Net for Epilepsy Prediction via EEG (Gram Residual Co-attention Net)

The GRC-Net for epilepsy prediction operates on transformed EEG time-series data (You et al., 13 Dec 2025). The raw 1D signal is first mapped to a 2D Gramian angular field (GAF), a representation encoding temporal correlations as an image, thus enabling spatio-temporal context modeling and improved noise suppression.

The architecture interleaves Inception-style residual units (enabling multi-scale local feature extraction) with a residual co-attention (CoT) module that captures global dependencies in the transformed EEG image. The CoT module computes a static local context key, fuses it with the input signal, projects the result through a series of convolutions, and mixes local/global information spatially via a lightweight attention-like mechanism with residual connections.

Following pooling, a fully connected layer produces class logits for seizure prediction. On the challenging five-class Bonn dataset, GRC-Net attains $93.66\%$ accuracy (F1: 93.14\%), outperforming prior CNN, LSTM, and ensemble approaches by a margin of up to 7 percentage points (You et al., 13 Dec 2025). Ablation studies validate the necessity of the GAF representation, the co-attention block, and the inception-style local modules for performance.

5. Cross-Model Comparison and Key Characteristics

GRC-Net Variant	Core Mechanism	Application Domain
Geometry-Reflectance Collab	Dual-modality encoding, info bottleneck, fusion	LiDAR semantic segmentation
Graph-Refined/Graph-Revised	Graph revision, prototype-routing, soft pruning	Recommendation, node classification
Global Receptive Conv (FCN+)	Local-global convolution channel split	Semantic segmentation
Gram Residual Co-attention	GAF encoding, inception+co-attention modules	EEG-based epilepsy prediction

The GRC-Net design pattern, across its instances, is characterized by:

• Early separation of heterogeneous or multi-modal inputs, often with distinct processing pipelines.
• Explicit mechanisms for information fusion, typically incorporating uncertainty weighting, cross-attention, or prototype-driven metric learning.
• Use of bottlenecks, global context modeling, or graph revision to enable robustness to input noise, domain shift, or structural imperfections.
• Minimal increase in model complexity relative to baseline, with attention to computational efficiency and scalability.

6. Impact, Limitations, and Research Directions

GRC-Net variants have established state-of-the-art performance in their respective areas, especially under conditions where conventional models are degraded by domain shift, label sparsity, or context deficiency. The explicit separation and recombination of heterogeneous features, as well as adaptive graph or signal revision, have shown to be vital for robustness in both vision and graph domains (Yang et al., 3 Jun 2025, Yinwei et al., 2021, Yu et al., 2019, Ren et al., 2023, You et al., 13 Dec 2025).

Limitations reported include reliance on fixed fusion or grouping parameters (e.g., grid partition in GRC blocks), potential overfitting in the absence of regularization for learned graph structure, and the challenge of extending revision/fusion mechanisms to more diverse graph or sequence modalities. Future work is likely to focus on adaptive or content-driven grouping/fusion, extension to directed or heterogeneous graphs, and broader application to regression and prediction tasks beyond classification.