MultiConv: Multi-kernel Gated Convolution

• MultiConv is a design pattern that combines multiple convolution kernels with adaptive gating to extract features at various scales.
• It reconciles heterogeneous local and global cues by modulating contributions of each branch, enhancing performance in tasks like imaging, video, and audio analysis.
• Empirical studies show that MultiConv significantly improves accuracy and robustness across applications, evidenced by gains in segmentation, classification, and detection metrics.

Searching arXiv for recent and foundational papers on Multi-kernel Gated Convolution and closely related formulations. Multi-kernel Gated Convolution, often abbreviated as MultiConv, denotes a family of architectural patterns in which multiple convolutional branches with different receptive fields are combined with a learned gating mechanism that modulates their contribution to the output. Across the literature, the pattern appears in decoder adapters for thyroid ultrasound, spatial aggregation blocks in modern ConvNets, recurrent ConvLSTM transitions for video, audio back-end classifiers built on SSL features, and DenseNet-like image classifiers with hybrid connectivity. Despite this diversity, the underlying objective is consistent: to reconcile feature cues that occur at different spatial or temporal scales while suppressing redundant, artifact-prone, or otherwise uninformative activations through gating (Sabouri et al., 9 Mar 2026, Li et al., 2022, Agethen et al., 2019, Tran et al., 3 Sep 2025, Yang et al., 2019).

1. Terminology and conceptual scope

The nomenclature is not standardized. In thyroid ultrasound, the paper does not define a separate module literally named “MultiConv”; rather, the Multi-Kernel Gated Adapter (MKGA) and ResMKGA instantiate a multi-kernel gated convolution pattern in the decoder (Sabouri et al., 9 Mar 2026). In MogaNet, the corresponding mechanism is called multi-order gated aggregation, housed in the spatial aggregation block $\mathrm{Moga}(\cdot)$ (Li et al., 2022). In ConvLSTM-based video modeling, the relevant terms are multi-kernel approach, Inception-like multi-kernel, and Network-in-LSTM (Agethen et al., 2019). In HCGNet, the relevant component is the multi-kernel depthwise convolution stage inside the SMG module, paired with update and forget gates (Yang et al., 2019). By contrast, the audio deepfake detector explicitly uses Multi-kernel gated Convolution (MultiConv) as the name of its back-end classifier block (Tran et al., 3 Sep 2025).

This variation in terminology reflects a broader fact: MultiConv is better understood as a design pattern than as a single canonical module. Some realizations emphasize spatially varying gates, some apply channel-wise softmax competition between branches, some use GLU-like multiplicative gating after branch fusion, and some embed the mechanism inside recurrent state transitions rather than feed-forward feature hierarchies. The common denominator is the joint use of multi-receptive-field convolutions and adaptive gating.

2. Canonical computational pattern

A general formulation appears in the thyroid ultrasound work, where the pattern is written for $m$ branches as

$$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$

followed by a projection

$X_{\text{ref}} = \phi_{1\times1}(Y),$

a context-dependent gate

$$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$

and either branch-wise gated fusion

$y_{\text{fused}} = \sum_i g_i \odot y_i$

or post-projection gating

$X_{\text{gate}} = g \odot X_{\text{ref}},$

before the final refinement stage (Sabouri et al., 9 Mar 2026). This is the clearest explicit abstraction of MultiConv in the cited literature.

The same logic recurs in specialized forms. In MogaNet, the multi-order context $Y_C$ is built from low-, middle-, and high-order branches, then projected and gated pointwise: $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$ (Li et al., 2022). In the ConvLSTM setting, the single-kernel input-to-hidden and hidden-to-hidden convolutions of each gate are replaced by aggregated multi-kernel transforms $W^{(mk)}_{xg} * X_t$ and $m$0, optionally preceded by kernel-specific attention masks derived from optical flow (Agethen et al., 2019). In the audio detector, MultiConv is realized by splitting the expanded representation into a left content stream and a right convolutional stream, fusing the multi-kernel outputs into $m$1, and computing

$m$2

as a GLU-like gate (Tran et al., 3 Sep 2025). In HCGNet, the gated fusion appears as

$m$3

where $m$4 is a forget-gate vector on reused features and $m$5 is a global context vector produced by update-gated fusion of the multi-kernel branches (Yang et al., 2019).

These formulations differ in granularity and parameterization, but they share three structural primitives: parallel convolutional context extraction, learned gating, and context-sensitive fusion.

3. Major architectural instantiations

Setting	Multi-kernel branches	Gating and fusion
Thyroid ultrasound MKGA/ResMKGA	$m$6 and dilated $m$7 with $m$8	Additive attention gate from $m$9 and $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$0; gated skip fused with decoder feature (Sabouri et al., 9 Mar 2026)
MogaNet spatial block	DW $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$1, DW $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$2, DW $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$3, with channel split $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$4	$$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$5 Conv + SiLU gate and context projection; element-wise product and residual addition (Li et al., 2022)
ConvLSTM video model	Parallel $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$6, $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$7, optionally $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$8	Concatenation with interleaving or learned $$y_i = k_i * x,\qquad i=1,\ldots,m,\qquad Y = [y_1;\ldots;y_m],$$9 mixing; optional kernel-specific attention masks from optical flow (Agethen et al., 2019)
Audio deepfake detector	Parallel 1D convolutions with kernel sets such as $X_{\text{ref}} = \phi_{1\times1}(Y),$0, $X_{\text{ref}} = \phi_{1\times1}(Y),$1, $X_{\text{ref}} = \phi_{1\times1}(Y),$2, $X_{\text{ref}} = \phi_{1\times1}(Y),$3, or four-kernel sets	Fused multi-kernel stream multiplicatively gates the split content stream; stacked across $X_{\text{ref}} = \phi_{1\times1}(Y),$4 layers (Tran et al., 3 Sep 2025)
HCGNet SMG module	Depthwise $X_{\text{ref}} = \phi_{1\times1}(Y),$5 and $X_{\text{ref}} = \phi_{1\times1}(Y),$6, or $X_{\text{ref}} = \phi_{1\times1}(Y),$7 dilated with $X_{\text{ref}} = \phi_{1\times1}(Y),$8 approximating $X_{\text{ref}} = \phi_{1\times1}(Y),$9	Update gate performs per-channel two-branch softmax after spatial attention and global pooling; forget gate decays reused features (Yang et al., 2019)

In the thyroid setting, MultiConv is explicitly decoder-side rather than backbone-side. MKGA refines skip features with complementary receptive fields and then gates them using semantic context from the deeper decoder state $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$0. ResMKGA adds residual bottleneck stabilization through squeeze-and-excitation, with

$$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$1

(Sabouri et al., 9 Mar 2026). This placement is motivated by cross-center domain shift, where segmentation relies more on geometry and malignancy prediction depends more on texture.

In MogaNet, MultiConv is part of the SMixer of a modern ConvNet block. It is preceded by the Feature Decomposition module

$$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$2

and followed by gated aggregation. The design is explicitly framed as a remedy for a representation bottleneck in modern ConvNets, particularly the under-encoding of middle-order interactions (Li et al., 2022).

In recurrent video modeling, MultiConv is embedded inside the ConvLSTM transition itself. Rather than using a single convolution for the gate preactivations, each recurrent gate receives a multi-branch transform, optionally deepened by $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$3 bottlenecks and mixing layers. This is why the authors describe the method as a Network-in-LSTM (Agethen et al., 2019).

In the audio model, the module is purely temporal and 1D. Its MultiConv branches operate on frame-wise SSL features from XLS-R after front-end SwiGLU gating and layer aggregation. The module is stacked, and diversity across layers is further promoted by a CKA term in the training objective (Tran et al., 3 Sep 2025).

HCGNet places the mechanism inside a two-stage SMG module: hierarchical squeezing first compresses DenseNet-like concatenated inputs, after which multi-kernel depthwise excitation and gated fusion produce a compact global-context update. Here MultiConv is tightly coupled to the network’s Hybrid Connectivity, which nests global dense and local residual pathways (Yang et al., 2019).

4. Empirical evidence across tasks and modalities

The thyroid ultrasound study provides a particularly direct demonstration of MultiConv under domain shift. With a ResNet34 backbone, the unfrozen baseline drops from in-domain Dice $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$4 to external Dice $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$5. Adding MKGA raises external Dice to $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$6, and ResMKGA raises it to $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$7, reported as significant at $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$8 versus the unfrozen baseline. For TI-RADS malignancy prediction, the same CNN baseline attains external AUC $$c = h(X_{\text{high}}, X_{\text{ref}}),\qquad g_i = \phi_{g,i}(W_{g,i}\cdot c + b_{g,i}),$$9, whereas MKGA reaches AUC $y_{\text{fused}} = \sum_i g_i \odot y_i$0 and improves diagnostic accuracy from $y_{\text{fused}} = \sum_i g_i \odot y_i$1 to $y_{\text{fused}} = \sum_i g_i \odot y_i$2 with McNemar $y_{\text{fused}} = \sum_i g_i \odot y_i$3 (Sabouri et al., 9 Mar 2026).

In MogaNet, the multi-order gated aggregation block scales from lightweight to very large regimes while remaining competitive with both ConvNet and ViT baselines. MogaNet-T reports $y_{\text{fused}} = \sum_i g_i \odot y_i$4 top-1 accuracy at $y_{\text{fused}} = \sum_i g_i \odot y_i$5 with $y_{\text{fused}} = \sum_i g_i \odot y_i$6M parameters and $y_{\text{fused}} = \sum_i g_i \odot y_i$7G FLOPs, and $y_{\text{fused}} = \sum_i g_i \odot y_i$8 at $y_{\text{fused}} = \sum_i g_i \odot y_i$9 with refined training. At the upper end, MogaNet-XL reports $X_{\text{gate}} = g \odot X_{\text{ref}},$0 at $X_{\text{gate}} = g \odot X_{\text{ref}},$1 with IN-21K pretraining, $X_{\text{gate}} = g \odot X_{\text{ref}},$2M parameters, and $X_{\text{gate}} = g \odot X_{\text{ref}},$3G FLOPs. The same design transfers to COCO detection, ADE20K segmentation, pose estimation, and video prediction (Li et al., 2022).

In ConvLSTM-based video recognition, replacing single-kernel recurrent convolutions with MultiConv improves performance on multiple settings. On UCF-101 with a VGG-16 feature extractor, the $X_{\text{gate}} = g \odot X_{\text{ref}},$4 baseline reaches $X_{\text{gate}} = g \odot X_{\text{ref}},$5 top-1, the $X_{\text{gate}} = g \odot X_{\text{ref}},$6 baseline $X_{\text{gate}} = g \odot X_{\text{ref}},$7, MultiConv with $X_{\text{gate}} = g \odot X_{\text{ref}},$8 reaches $X_{\text{gate}} = g \odot X_{\text{ref}},$9, and adding stacked $Y_C$0 mixing yields $Y_C$1. On Sports-1M-20, the corresponding values are $Y_C$2, $Y_C$3, and $Y_C$4. In an I3D-based UCF-101 setting, Inception MultiConv reaches $Y_C$5, and the end-to-end version reaches $Y_C$6 (Agethen et al., 2019).

The audio deepfake detector reports state-of-the-art or competitive error rates on multiple benchmarks. It attains $Y_C$7 EER on ASVspoof 2019 LA, $Y_C$8 on ASVspoof 2021 DF, $Y_C$9 on ASVspoof 2021 LA, and $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$0 on ITW. The four-kernel configuration $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$1 with CKA is reported as best for heterogeneous out-of-domain generalization, and removing CKA degrades performance, for example from $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$2 to $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$3 on 21DF and from $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$4 to $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$5 on DFADD in the cited comparison (Tran et al., 3 Sep 2025).

HCGNet provides evidence for MultiConv in dense image classification and transfer. On ImageNet, HCGNet-B reports top-1/top-5 error $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$6 with $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$7M parameters and $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$8G FLOPs, compared with DenseNet-169 at $$g(U)=\mathrm{SiLU}(W_gU),\qquad a(U)=\mathrm{SiLU}(W_cY_C),\qquad y=g(U)\odot a(U),\qquad Z=X+y$$9, $W^{(mk)}_{xg} * X_t$0M parameters, and $W^{(mk)}_{xg} * X_t$1G FLOPs. On CIFAR-10/100, HCGNet-A2 reaches $W^{(mk)}_{xg} * X_t$2 and $W^{(mk)}_{xg} * X_t$3 error with $W^{(mk)}_{xg} * X_t$4M parameters and $W^{(mk)}_{xg} * X_t$5G FLOPs, while HCGNet-A3 reaches $W^{(mk)}_{xg} * X_t$6 and $W^{(mk)}_{xg} * X_t$7 with $W^{(mk)}_{xg} * X_t$8M parameters and $W^{(mk)}_{xg} * X_t$9G FLOPs (Yang et al., 2019).

Taken together, these results show that MultiConv is not confined to one task family. It has been used for segmentation, image classification, action recognition, audio spoofing detection, pose estimation, and detection/segmentation transfer, with gains emerging when the target problem contains heterogeneous local and global cues.

5. Ablations, trade-offs, and failure modes

The thyroid work offers one of the clearest ablation studies on the role of gating. Removing the gate slightly improves external segmentation, with Dice $m$00 instead of $m$01, but sharply degrades TI-RADS classification, with accuracy $m$02 instead of $m$03 and AUC $m$04 instead of $m$05. Removing the multi-kernel block harms both tasks, producing DDTI Dice $m$06, AUC $m$07, and accuracy $m$08. Among kernel choices, $m$09 plus dilated $m$10 with effective $m$11 receptive field is reported as the best balance, whereas $m$12 loses diagnostic power and $m$13 is unstable (Sabouri et al., 9 Mar 2026).

MogaNet reports a similar dependence on branch design and gating details. In the module ablation table, adding the gating branch raises accuracy from $m$14 to $m$15, adding DW $m$16 raises it to $m$17, adding multi-order DW convolutions raises it to $m$18, adding FD raises it to $m$19, and replacing SE/MLP with the CA block yields $m$20. Removing $m$21, removing $m$22, or removing multi-order DWConv each lowers top-1 accuracy, and SiLU in both branches is reported as the best gating activation. The default channel split $m$23 is also identified as the best among tested splits (Li et al., 2022).

In ConvLSTM, the benefits of MultiConv depend on motion heterogeneity. Mixed channel allocations across $m$24 and $m$25 kernels outperform pure single-kernel settings, and qualitative results show specialization of attention masks: larger kernels attend to faster-moving digits, smaller kernels to slower motion. Yet the paper also notes that when motion speeds are relatively uniform and fit within the receptive field of a small kernel, multi-kernel mixtures may provide limited gains or even underperform a well-matched single kernel (Agethen et al., 2019).

The audio detector reports that no single kernel configuration universally dominates. Smaller kernels $m$26 are best on ITW and LibriSeVoc, large kernels $m$27 excel on 19LA and HABLA, and the four-kernel set $m$28 generalizes best to heterogeneous out-of-domain attacks. The model also struggles with neural autoregressive vocoders and certain TTS systems, and Sino-Tibetan datasets remain difficult. Deeper stacks beyond the default $m$29 can improve in-domain performance but hurt generalization (Tran et al., 3 Sep 2025).

HCGNet identifies additional constraints. Its MultiConv stage uses only two scales, and descriptor-level fusion may underweight fine-grained spatial patterns because the update gate aggregates global descriptors rather than spatially varying fusion maps. The paper also notes the usual depthwise-convolution trade-off: efficiency is improved, but cross-channel mixing becomes weaker and must be supplied by the squeeze path and gating FCs (Yang et al., 2019).

These studies collectively show that MultiConv is not a monotone recipe in which “more kernels” or “stronger gates” always improve outcomes. Kernel size, branch count, gating granularity, and the statistics of the target data interact strongly.

6. Relation to adjacent methods and likely directions

Several papers explicitly position MultiConv relative to adjacent design families. MogaNet distinguishes its multi-order gated aggregation from Inception-like and SKNet-style branch selection: it constructs a unified multi-order context with parallel depthwise kernels and dilations, then applies a pointwise SiLU gate to the aggregated context rather than branch-wise softmax competition (Li et al., 2022). HCGNet similarly distinguishes its SMG module from Inception-style concatenation and from SKNet’s direct kernel selection on branch feature maps, emphasizing instead spatial attention, global descriptor extraction, and per-channel two-branch softmax fusion at the descriptor level (Yang et al., 2019). The thyroid study compares MKGA against raw skip fusion, single-kernel conv fusion, SE-only recalibration, and attention-only fusion, arguing that robustness under cross-center shift depends on the combination of complementary receptive fields and semantic gating before decoder fusion (Sabouri et al., 9 Mar 2026).

This suggests that MultiConv occupies an intermediate position between static multi-branch convolution and attention-like dynamic modulation. It preserves the locality and inductive bias of convolutions while introducing data-dependent selection mechanisms that are lighter than full self-attention in several of the cited settings.

The forward directions mentioned in the literature are concrete. In thyroid ultrasound, the single post-projection spatial gate could be replaced by branch-wise or channel-wise gating, and extension to 3D ultrasound would require volumetric MultiConv and 3D gating; neither is explored in the paper (Sabouri et al., 9 Mar 2026). In ConvLSTM, the attention mechanism is applied only to the input-to-hidden path, with hidden-to-hidden attention reserved for future work (Agethen et al., 2019). In the audio model, kernel-set selection remains task-dependent, and the authors’ results imply that tuning receptive-field ensembles to attack families and languages is important (Tran et al., 3 Sep 2025).

Across these variants, the enduring contribution of MultiConv is not a single invariant block definition, but a reusable architectural principle: parallel receptive fields are most effective when paired with an explicit mechanism for deciding what information should pass through them, where, and at what scale.