Multi-Branch 1D CNNs

• Multi-branch 1D CNNs are neural architectures that use independent 1D convolution branches to capture multi-scale or axis-specific features in structured data.
• They fuse the outputs of separate branches through concatenation, enhancing robustness and accuracy in tasks such as mesh segmentation and acoustic scene classification.
• These architectures offer improved computational efficiency and interpretability by isolating feature channels and reducing redundant spatial correlations.

Multi-branch 1D Convolutional Neural Networks (CNNs) are neural architectures in which multiple independent branches, each composed of one-dimensional convolutional layers, operate in parallel on feature representations. Each branch typically learns to extract complementary features from distinct subspaces or at different scales, augmenting the expressive capacity of the architecture over conventional single-branch or stacked 1D convolutional designs. Branch outputs are ultimately fused before downstream prediction layers, supporting more discriminative and robust representations. Two notable instantiations are the multi-scale mesh segmentation architecture of George et al. (George et al., 2017) and the axis-separating TF-SepNet for acoustic scene classification (Cai et al., 2023).

1. Mathematical Formalism and 1D Convolutional Building Blocks

A 1D convolutional layer in multi-branch designs applies a discrete filter $f$ of length $K$ to an input feature $x$ using zero-padding as follows: $$(f * x)[i] = \sum_{j=-\lfloor K/2\rfloor}^{\lfloor K/2\rfloor} f[j] \; x[i-j]$$ This operation is applied channel-wise, and results may be summed over channels. In both mesh and audio applications, 1D convolutions are favored over reshaped 2D alternatives to preserve the true semantic relationships in sequential feature inputs and eliminate spurious spatial correlations.

Distinct branches may employ different convolution parameters—such as kernel sizes ($K=15$ or $K=11$ in (George et al., 2017)), stride, or depthwise grouping—as well as varied pre- and post-processing (e.g., batch normalization, nonlinearities, pooling, or global pooling operations).

2. Multi-Scale and Multi-Axis Branch Design

Mesh Segmentation (Multi-Scale)

In the segmentation network for 3D meshes (George et al., 2017), three branches operate at independent geodesic radii, accepting as input feature vectors aggregated over face, 1-ring, and 2-ring neighborhoods. Each branch processes its input via two 1D convolutional and max-pooling stages with independent parameters:

• Conv1: 16 filters, $15\times 1$ kernels, Leaky ReLU
• Pool1: $2\times$ downsampling
• Conv2: 32 filters, $11\times 1$ kernels
• Pool2: $2\times$ downsampling

Acoustic Scene Classification (Multi-Axis)

TF-SepNet (Cai et al., 2023) utilizes parallel branches explicitly aligned with the time and frequency axes, operating on audio spectrograms. The repeated "TF-SepConvs" module splits the incoming feature into two channel blocks:

• The "frequential path" applies a $K$0 depthwise convolution and average-pools along the frequency axis, then uses a $K$1 pointwise convolution.
• The "temporal path" applies a $K$2 depthwise convolution and average-pools in the time dimension, then again uses a $K$3 convolution.

Both strategies exploit branch architectural independence to enable extraction of either multi-scale or axis-specific information, followed by channel-wise recombination.

3. Feature Construction and Branch Inputs

Mesh Networks

For each mesh face, input is a high-dimensional vector consisting of:

• 593 standard geometric descriptors (including curvatures, PCA, medial distance)
• 6 conformal factor features (including non-shrinking Laplacian smoothed variants)
• Heat Kernel Signature samples (100 original, 100 scale-invariant) This 800-dimensional vector is further aggregated over neighborhood rings to produce per-branch inputs (George et al., 2017). No PCA or feature reduction precedes CNN processing in each branch.

Acoustic Networks

TF-SepNet operates on spectrogram-like representations $K$4, splitting features along the channel axis for dedicated time and frequency processing. Channel shuffling precedes branch division to maximize feature diversity (Cai et al., 2023).

4. Branch Fusion and Network Topologies

Mesh Segmentation (George et al., 2017):

• Three independent branches yield $K$5 outputs each, concatenated to form $K$6 representations.
• Subsequent fully connected layers reduce this to class prediction logits, with nonlinearity and dropout applied pre-output.

TF-SepNet (Cai et al., 2023):

• Each "TF-SepConvs" module produces outputs along time and frequency axes, which are broadcast and recombined via direct channel concatenation, preserving the full learned capacity from both branches.
• Subsequent stacking of modules, pointwise convolution, and global pooling deliver predictions.

No intermediate inter-branch communication or attention is used; fusion relies exclusively on concatenation, preserving information from each independent processing stream.

5. Effective Receptive Field and Computational Efficiency

TF-SepNet demonstrates that multi-branch 1D structures with global axis-wise pooling achieve effective receptive fields (ERF) spanning the entire input along the corresponding axis after a single module:

• Frequential: $K$7; Temporal: $K$8
• By contrast, consecutive 1D filters require $K$9 or $x$0 layers to reach full-receptive-context.

In mesh networks, multiple branches allow for hierarchical receptive fields over local to global mesh neighborhoods. Both designs deliver improved parameter and FLOP efficiency relative to conventional 2D convolutional stacks. For example, TF-SepNet-40 attains 53.4K parameters and 7M MACs per inference—substantially lower than comparable 2D and consecutive-1D baseline designs (Cai et al., 2023).

6. Quantitative Performance and Ablation Studies

Mesh Segmentation (George et al., 2017):

• On the Princeton Segmentation Benchmark, multi-branch 1D CNN achieves leave-one-out mean accuracy of 94.80%, surpassing the best 2D CNN's 92.79%.
• Adding branches consistently improves accuracy: 1 branch (93.57%), 2 branches (94.22%), 3 branches (94.63%), 4 branches (94.81%).
• Using only legacy features, the architecture retains superiority over prior 2D-based approaches.

Acoustic Scene Classification (Cai et al., 2023):

• On the TAU Urban Acoustic Scene 2022 Mobile dataset, TF-SepNet-40 achieves 60.0% top-1 accuracy with significantly reduced MACs and parameter count relative to BC-ResNet and BC-Res2Net baselines.
• Ablation reveals 2–3% accuracy drop when either branch (time or frequency) is removed, and further shows the necessity of both channel shuffling and adaptive normalization for peak performance.
• ERF analysis: high-contribution area (threshold $x$1) ratio $x$2 increases from 39.3% (baseline) to 43.8% (TF-SepNet-40), indicating more uniform spatial influence.

7. Advantages, Limitations, and Prospective Extensions

Advantages:

• Multi-branch 1D CNNs prevent contamination of feature channels by arbitrary reshaping, yield higher interpretability, and reflect true problem geometry (mesh rings, axes in audio) (George et al., 2017, Cai et al., 2023).
• Branch independence enables more effective multi-scale/axis learning and mitigates overfitting to local feature noise.
• Computational and memory efficiency is substantively improved through use of depthwise or grouped convolutions and parallelized topologies.

Limitations:

• The mesh segmentation approach depends on precomputed, hand-crafted features and fixed filter parameters; it is not end-to-end or directly extensible to raw geometric data (George et al., 2017).
• TF-SepNet's global pooling in each branch, while conferring large ERF, may obscure fine-grained positional variance.

Extensions:

• Adoption of learned per-vertex embeddings, graph convolutional structures, or integration of residual-inception modules could further enhance representational depth (George et al., 2017).
• Design variants with attentive/inter-branch fusion, synthetic augmentation, or multi-view consistency losses are plausible future directions.
• Both paradigms provide templates for other domains in which data exhibits natural multi-scale or multi-axis structure.

---

Multi-branch 1D CNN architectures provide a principled, efficient, and empirically validated approach for structured signal processing in domains ranging from geometric mesh analysis to time–frequency audio understanding. By leveraging independent 1D convolutional branches with targeted receptive fields, these models support improved accuracy, robustness, and computational scalability compared to both 2D and single-branch 1D convolutional approaches (George et al., 2017, Cai et al., 2023).