Separable Group Convolutional Framework

• Separable Group Convolutional Framework is a method that factorizes kernels into distinct spatial, subgroup, and channel components to exploit redundancy and maintain group equivariance.
• The framework leverages rigorous group theory and kernel decomposition techniques to significantly reduce parameters and computational cost in various applications.
• Empirical analyses demonstrate improved accuracy and efficiency on tasks like MNIST-rot and volumetric medical imaging, highlighting actionable design trade-offs in network architecture.

Searching arXiv for relevant papers on separable group convolutions and related frameworks. A separable group convolutional framework is a family of group-equivariant convolutional constructions in which the convolution kernel is factorized across spatial, subgroup, and sometimes channel dimensions in order to exploit redundancy while preserving the equivariance structure of the original operator. In the Lie-group setting emphasized by separable G-CNNs, a kernel on $G = \mathbb{R}^n \times H$ is written as $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$, so that convolution along the subgroup $H$ and convolution over the spatial domain can be treated as distinct operations; related work extends the same principle to channelwise factorization, continuous kernel parameterization on Lie algebras, and separable SE(3) kernels for volumetric data (Knigge et al., 2021, Lengyel et al., 2021, Kuipers et al., 2023).

1. Precursors in separable and grouped convolution

In standard CNNs, separable convolution is mathematically defined as a depthwise convolution followed by a pointwise $1\times1$ convolution. One interpretation describes the depthwise stage as a grouped convolution with maximum cardinality, while the $1\times1$ stage plays an Inception-like role by mixing features across channels (Ghosh, 2017). This grouped-plus-pointwise motif is central to later separable group convolutional designs.

A related precursor is the decomposition of a $K\times K$ spatial filter into a linear combination of rank-1 filters,

$W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$

implemented as a three-stage block consisting of horizontal $K\times1$ convolutions, vertical $1\times K$ grouped convolutions, and $1\times1$ fusing convolutions. In that setting, the complexity per layer drops from $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$0 to $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$1, with a reported $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$2 reduction in the number of multiplications and about $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$3 overall speedup on an industrial digit/letter recognition task, while preserving accuracy when the $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$4 fusing stage is retained (Limonova et al., 2020). In sequence modeling, depthwise separable convolutions and the related super-separable convolution similarly reduce parameters from $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$5 to $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$6, or to $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$7 under grouped pointwise mixing, and enable wider convolution windows without dilation in SliceNet (Kaiser et al., 2017).

A plausible implication is that separable group convolution inherits not just a computational trick, but a general architectural principle: isolate a structured transformation axis, apply a low-cost operation along that axis, and reintroduce coupling through explicit mixing.

2. Redundancy in group convolution kernels

The immediate motivation for separable group convolutions is empirical redundancy in regular G-CNN filters. In a standard group convolution, the input tensor has shape $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$8, the filter tensor has shape $k(x,h) = k_{\mathbb{R}^n}(x)\cdot k_H(h)$9, and the output is computed as

$H$0

Analyses of trained models showed that filters frequently become highly redundant along the group axis: if kernels in that dimension are identical or rotationally symmetric, feature maps become identical along the group dimension, inducing identical gradients in the next layer; PCA of learned filters confirms that the first principal component can explain a large portion of the variance across the group dimension (Lengyel et al., 2021, Knigge et al., 2021).

This observation leads to explicit factorization. A first decomposition writes

$H$1

where $H$2 is a shared spatial kernel and $H$3 contains group-dependent coefficients. A stronger variant additionally shares the spatial kernel across input channels,

$H$4

These parameterizations yield an efficient implementation as a group-wise pointwise operation followed by a spatial group-transformed convolution, while preserving equivariance because the full transformed filter bank is still generated through the group action $H$5 (Lengyel et al., 2021).

The parameter savings are explicit. Standard GConv uses $H$6 parameters, the $H$7-GConv parameterization uses $H$8, and the stronger $H$9-GConv uses $1\times1$0 (Lengyel et al., 2021). This suggests that the principal benefit of the framework is not merely fewer coefficients, but weight tying over degrees of freedom that regular training already tends to collapse.

3. Kernel factorization on Lie groups

For affine Lie groups of the form $1\times1$1, separable G-CNNs factorize the group kernel as

$1\times1$2

The corresponding group convolution,

$1\times1$3

can then be rewritten so that the subgroup component is applied first,

$1\times1$4

followed by a spatial convolution with $1\times1$5 (Knigge et al., 2021). In this formulation, separability is not a post hoc approximation but the defining kernel parameterization.

The framework also supports channelwise separation. A channel-indexed kernel may be written as

$1\times1$6

and the empirically preferred configuration applies additional spatial-depthwise separation: the spatial kernel depends only on output channels and is shared across input channels (Knigge et al., 2021). The reported complexity reduction for the subgroup factorization is from

$1\times1$7

This computational reduction is what makes higher-order equivariance practical. In particular, the same framework was used to implement G-CNNs equivariant to $1\times1$8, the group of dilations, rotations and translations, which would otherwise be computationally prohibitive in a non-separable form (Knigge et al., 2021).

4. Continuous parameterization and discretization

A central issue for group-equivariant models is how to parameterize kernels on continuous groups. In separable G-CNNs on Lie groups, the kernel is implemented as a Sinusoidal Representation Network,

$1\times1$9

where $1\times1$0 maps a group element to its Lie algebra coordinate. For the separable kernel, $1\times1$1 and $1\times1$2 are implemented as separate SIRENs. Continuous group integrals are approximated either by discretization or by random sampling, and uniform random sampling is reported as an unbiased estimator of the group convolution integral that often gives superior results to fixed discretization (Knigge et al., 2021).

A related continuous formulation appears in regular SE(3) group convolutions for volumetric medical image analysis. There, the SE(3) kernel is separated into a continuous SO(3) kernel and a spatial kernel,

$1\times1$3

with the integral over SO(3) approximated by a finite sum over a uniform SO(3) grid. The continuous SO(3) kernel is parameterized by RBF interpolation,

$1\times1$4

where $1\times1$5 is the geodesic distance on SO(3) and $1\times1$6 is a Gaussian RBF (Kuipers et al., 2023).

The significance of these constructions is that they separate the question of equivariance from the question of storage. Continuous kernels are represented by compact parametric functions, then sampled only as needed for numerical integration or for generating transformed filter copies.

5. Representative instantiations and empirical behavior

Several empirical instantiations illustrate the range of the separable group convolutional framework.

Framework	Domain or group	Reported outcome
Separable G-CNNs on Lie groups	SE(2), $1\times1$7, $1\times1$8	$1\times1$9 on MNIST-rot; $K\times K$0 on CIFAR10; $K\times K$1 on CIFAR100; $K\times K$2 on Galaxy10
SepGrouPy-style separable GConvs	p4, p4m	gc-p4m-ResNet44: $K\times K$3 vanilla and $K\times K$4 with data augmentation; $K\times K$5M parameters vs. $K\times K$6M
Separable SE(3) group convolutions	volumetric medical image analysis	up to a $K\times K$7 gain in accuracy over regular CNNs

In the Lie-group framework, separable G-CNNs often outperform non-separable G-CNNs while using only a fraction of their training time, and $K\times K$8-equivariance further improves performance on all tasks considered. On MNIST-rot, the best reported test error is $K\times K$9 for a $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$0-CNN with train-time augmentation; on CIFAR10, the best overall accuracy reported is $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$1 for a SIREN-16 $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$2 model, with $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$3 listed in the summary table; CIFAR100 reaches $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$4, and Galaxy10 $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$5 (Knigge et al., 2021).

In the discrete p4 and p4m setting, separable GConvs are reported to outperform standard GConv architectures on Rotated MNIST and CIFAR10 for the same or fewer parameters. On Rotated MNIST, a baseline P4CNN obtains $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$6 test error with $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$7K parameters, while separable variants span $W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$8–$W \approx \sum_{i=1}^{r} a_i \cdot u_i v_i^T,$9 for g-P4CNN and $K\times1$0–$K\times1$1 for gc-P4CNN, depending on parameter budget. On CIFAR10, gc-p4m-ResNet44 achieves the lowest error at $K\times1$2 without augmentation and $K\times1$3 with augmentation, using $K\times1$4M parameters rather than $K\times1$5M (Lengyel et al., 2021).

For volumetric medical image analysis, separable SE(3)-equivariant models consistently outperform vanilla CNNs and regular discrete G-CNNs on OrganMNIST3D, FractureMNIST3D, NoduleMNIST3D, AdrenalMNIST3D, and SynapseMNIST3D. The reported gain reaches $K\times1$6 percentage points over CNNs, and the accuracy drop under random test-time rotation is markedly smaller for SE(3)-CNNs, with drops of $K\times1$7–$K\times1$8 compared with $K\times1$9 for regular CNNs (Kuipers et al., 2023).

6. Generalizations, trade-offs, and related directions

A broader generalization is given by Lie group algebra convolutional filters. In that framework, traditional group convolutions are treated as one particular instantiation of a Lie group algebra homomorphism

$1\times K$0

with filters in $1\times K$1 and signals in a Hilbert space $1\times K$2. The discretization of the filter and the discretization of the signal are explicitly decoupled: filters are sampled on the group via the exponential map and products of Lie algebra generators, while signals may be sampled on arbitrary spaces and transformed back to the sampling domain by interpolation. The framework proves uniqueness of representation for bandlimited filters and derives the operator error bound

$1\times K$3

linking reconstruction accuracy to sampling density and filter bandwidth (Kumar et al., 2023). This suggests a more general notion of separability in which filter-space structure and signal-space structure are independent design variables.

The framework also has clear limitations. In separable G-CNNs on Lie groups, non-separable models slightly outperform separable ones on Scaled MNIST, which is attributed to a need for distinct kernels per scale (Knigge et al., 2021). In SE(3)-CNNs, higher SO(3) resolution increases memory use and training time per epoch, and when the data is highly isotropic, strict equivariance can slightly hinder performance because “too much” equivariance may prevent modeling dataset-specific biases (Kuipers et al., 2023). In non-group separated-filter CNNs, omitting the $1\times K$4 fusing stage causes a clear accuracy drop: for digit recognition, the standard CNN error is $1\times K$5, the proposed CNN without fusing gives $1\times K$6, and the proposed CNN with fusing gives $1\times K$7; for letters, the corresponding numbers are $1\times K$8, $1\times K$9, and $1\times1$0 (Limonova et al., 2020).

A further design consequence is that separability often shifts, rather than eliminates, computational bottlenecks. In ResNet-50, after replacing $1\times1$1 convolutions with depthwise separable convolution, $1\times1$2 convolutions account for about $1\times1$3 of the compute; fully learnable group convolution was proposed precisely to accelerate this regime, yielding about $1\times1$4 theoretical reduction in MAdds and $1\times1$5 actual speedup in single-CPU inference time when combined with depthwise separable convolution, with accuracy almost fully retained (Wang et al., 2019). This underscores a recurrent pattern across separable group convolutional frameworks: once spatial or subgroup interactions are factorized, efficient and expressive channel or group mixing becomes the critical remaining operation.