ChannelNets: Efficient Channel-Wise Architectures

• ChannelNets are a family of architectures that explicitly model channel-wise operations and sparsity to reduce parameters and computational load.
• They employ techniques like group, channel-local, and depth-wise convolutions, alongside wavelet-based channel attention and balanced width search methods.
• ChannelNets enhance robustness and efficiency across diverse applications including CNNs, massive MIMO detection, RF fingerprinting, and decentralized cross-chain networks.

ChannelNets are a family of architectures and learning principles that leverage explicit modeling of channel-wise operations, dependencies, and sparsity for efficiency, representation learning, and robustness in diverse neural network and distributed systems contexts. This term has been applied in the literature to a spectrum of methods ranging from lightweight convolutional neural architectures using channel-local or channel-wise convolutions, to width search supernets, robust detection for massive MIMO, RF fingerprinting under unknown multipath channels, and decentralized cross-chain networks with the notion of "channels" in distributed ledgers.

1. Channel-wise Designs in Convolutional Neural Networks

ChannelNets in the context of convolutional architectures designate lightweight neural networks that systematically replace dense channel-mixing (as present in standard $1{\times}1$ convolutions or fully connected layers) with highly sparse channel-wise operations, yielding drastic reductions in both parameter count and computational load (Gao et al., 2018). Specifically:

• Group channel-wise convolution (GCWConv): Inputs and outputs are partitioned into $g$ channel groups and processed with $1{\times}1$ group convolutions followed by independent 1-D channel-wise convolutions, thus restoring cross-group mixing lost in standard group convolutions.
• Depth-wise separable channel-wise convolution (DWSCWConv): Replaces depth-wise separable blocks with a depth-wise 2D spatial convolution and a global 1-D channel-wise convolution, decoupling spatial from channel mixing.
• Convolutional classification layer (CCL): The dense final fully connected classifier is replaced with a small 3-D convolutional kernel along the channel axis, which leverages empirically observed classifier sparsity to retain accuracy with orders-of-magnitude fewer parameters.

These operations are stacked into full-featured CNNs (e.g., ChannelNet-v1/v2/v3) following a MobileNet-like macro-architecture. Empirically, these networks achieve on ImageNet-1K up to $28\%$ reduced FLOPs and $12\%$ fewer parameters for matched accuracy relative to MobileNet 1.0, and maintain strong accuracy even at extremal compactness (e.g., ChannelNet-v3 with only $1.7$M parameters) (Gao et al., 2018).

Efficiency Rationale

Original 2D convolution has parameter and FLOP complexity scaling as $d_k^2mn$ (kernel size, input and output channels). ChannelNets effectively exploit the observation that dense channel mixing is highly redundant—in measured classifiers, $>90\%$ of weights are near zero, motivating sparse channel-wise replacements. The approach also overcomes "group isolation" in classical group convs by using channel-wise fusions, which require only $O(gd_c)$ additional parameters per block ($g$ groups, channel filter size $d_c$).

2. Channel Local Convolutions and CLCNet

The generalization of channel-wise operations leads to the formal notion of channel local convolutions (CLC) (Zhang, 2017). In CLC, each output channel $y_j$ is computed using only a subset $S_j$ of the $M$ input channels, i.e.,

$y_j = \sum_{i\in S_j} w_{j,i} \star x_i,$

where $w_{j,i}$ is a spatial kernel and the dependency structure is encoded via a channel dependency graph (CDG).

Key special cases include:

• Regular convolution: Every output aggregates over all channels ($|S_j| = M$).
• Grouped convolution: Each output channel accesses only the channels in its group ($|S_j| = M/g$).
• Depthwise convolution: Each output depends on a single input channel.

CLCNet introduces interlaced grouped convolutions (IGC), which permute group assignments between successive layers to ensure full channel receptive field (FCRF), while maintaining sparsity. Optimization of group sizes $g_1, g_2$ per block (stacking IGC and GC) under the FCRF constraint $g_1g_2 \le L$ is used to minimize FLOPs as

$C(g_1, g_2) = A M L / g_1 + L N / g_2,$

with $A = K^2$ for $K{\times}K$ spatial kernels.

CLCNet achieves superior accuracy-parameter-resource balance versus MobileNet and ShuffleNet, with clcNet-B outperforming MobileNet (Top-1 accuracy $71.6\%$ vs $70.6\%$, $25\%$ fewer mult-adds) and benefiting from significantly lower inference latency on mobile hardware (Zhang, 2017).

3. Channel Width Search and Bilaterally Coupled Networks

Channel-wise optimization extends beyond efficient design to network width search. The Bilaterally Coupled Network (BCNet) paradigm addresses training fairness in supernet-based width search, where traditional unilaterally augmented (UA) supernets suffer from channel selection bias—earlier channels are present in exponentially more subnetworks (Su et al., 2022). BCNet mitigates this by constructing a paired supernet that symmetrically selects both leftmost and rightmost channel subsets for each width, ensuring that every channel is trained equally often.

An enhanced method, BCNetV2, maintains fairness within variable-width ranges by overlapping the two supernets and employs alternating supernet forward passes for memory efficiency. Complementary stochastic sampling further corrects for finite-sample imbalance. The accompanying benchmark Channel-Bench-Macro provides exhaustive accuracy, parameter, and FLOPs data for $32,768$ architectures on CIFAR-10, enabling reproducible ranking of width search algorithms.

This methodology improves both pruning/search accuracy and deployment performance, with, for example, +$0.65\%$ absolute Top-1 improvement on ImageNet by refining EfficientNet-B0 under fixed FLOPs.

4. ChannelNets in Massive MIMO Detection

The ChannelNet architecture described in the context of massive MIMO detection (Ye et al., 2024) is a purely data-driven DNN detector in which the channel matrix $\mathbf H$ is not treated as part of the input but is embedded as fixed linear layers within the network. Each iteration (layer) consists of:

• Antenna-wise MLP updates for both receive and transmit features,
• Exchange via $\mathbf H$ and $\mathbf H^{T}$ between antenna domains,
• Iterative refinement with deep skip connections.

ChannelNet achieves computational complexity $O(N_t N_r)$ per detection (with $N_t, N_r$ antennas), matching approximate message passing (AMP) but with superior robustness and performance, including under correlated channels, imperfect CSI, or non-Gaussian noise. A key theoretical result is that ChannelNet-MLP is a universal approximator in probability for continuous permutation-equivariant functions of the channel and received signal, which includes the ML-optimal detector as a special case (Ye et al., 2024).

5. Channel-Wise and Group Equivariance in RF Fingerprinting

In the context of RF fingerprinting under multipath channel uncertainty, ChaRRNet introduces complex-valued group-convolutional layers that are equivariant to the Lie group of finite impulse response (FIR) channels (Brown et al., 2021). The layers operate on a feature space indexed by both channel and time, with kernels tied across group translates to ensure group-equivariance. Pooling over group elements yields channel-invariant representations, enabling robust device identification under domain shift (e.g., different physical environments).

This approach achieves significant accuracy gains on both synthetic and real-world datasets when trained and tested under mismatched channel conditions, demonstrating that explicit channel-group modeling and invariance are powerful inductive biases for wireless signal processing.

6. Channel Attention: Wavelet Compression and ChannelNet Generalization

“ChannelNet” also appears as a term describing a generalization of channel-attention mechanisms, specifically, replacing the scalar global average pooling (GAP) used in Squeeze-and-Excitation (SE) networks with a multilevel wavelet transform (Salman et al., 2022). WaveNet replaces the GAP "squeeze" step with recursive Haar (or general orthonormal) wavelet-based compression, rigorously shown to be equivalent to GAP when using the lowest-frequency subband but yielding richer representations when additional subbands are included.

Ablation studies reveal that introducing custom orthonormal filters and mixing additional bands produces incremental gains in classification accuracy with negligible overhead, pointing to multi-scale channel compression as a superior approach for information-preserving channel attention.

7. Channel Networks in Decentralized Systems

ChannelNet has also been adopted in a distributed systems context as the Cross-Chain Channel Network (CCN), a decentralized framework enabling secure and privacy-preserving multi-hop cross-chain interactions (Xu et al., 3 Dec 2025). Here, “channels” refer to off-chain payment channels or smart contract-based atomic transfer routes. The R-HTLC protocol within CCN uses hash-locks, zk-SNARKs, and a novel hourglass mechanism to guarantee atomic settlement and unlinkability, resolving both "active" and "passive" offline adversarial cases. CCN achieves atomicity, computational practicality, and privacy while scaling across heterogeneous blockchain platforms.

Conclusion

The ChannelNet/ChannelNets terminology encompasses a continuum of approaches uniting explicit channel-wise computations, sparsity, group-theoretic invariance, width search, and privacy-preserving multi-hop interaction. These developments are characterized by efficient parameter usage, theoretical approximation universality, robustness under domain shift or adversarial conditions, and empirical performance parity or superiority over dense or baseline architectures across computer vision, communications, and distributed ledger domains (Gao et al., 2018, Zhang, 2017, Su et al., 2022, Ye et al., 2024, Brown et al., 2021, Salman et al., 2022, Xu et al., 3 Dec 2025).