MCL: Meta Convolution Layer

Naim Reza; Md Al Amin; Ho Yub Jung

cs.CVOct 8, 2026

MCL: Meta Convolution Layer

Authors: Naim Reza, Md Al Amin, Ho Yub Jung

Organizations: Department of Computer Engineering, Chosun University, Gwangju 61452, Republic of Korea

Abstract

Dynamic convolution enhances convolutional neural networks (CNNs) by adapting kernels to input content, but it expresses the effective kernel as a linear mixture of a small number of basis kernels, which limits expressivity and complicates optimization as the mixture size grows. In this work, we revisit dynamic convolution from a functional perspective and propose the Meta Convolution Layer (MCL), which directly models the convolutional kernel as an input-conditioned function W(x) realized via a high-order polynomial expansion. Leveraging nested residual blocks inspired by deep polynomial networks, MCL implements a structured polynomial meta-network that generates a single input-adaptive kernel, thereby decoupling representational power from the explicit number of mixture kernels and alleviating training instability. MCL is a plug-in addition with standard convolutions and can be seamlessly integrated into both CNN and transformer backbones. Experimental evaluation shows that adding MCL improves the Top-1 accuracy of Resnet- 18, Resnet-50 and ResNet-101 by 6.61%, 3.42% and 3.05% on the ImageNet dataset. Moreover, the proposed method significantly boosts the accuracy of Resnet and Wide-Resnet variants on CIFAR-10 and CIFAR-100 datasets. Additionally, the proposed method outperforms previous methods on fine-grained visual classification tasks using Swin and ViT backbones. These results demonstrate that high-order polynomial kernel generation is a powerful and scalable alternative to linear mixture based dynamic convolution.

Figures & tables

Fig. 1: Schematic diagram of the proposed MCL method.

Fig. 2: Schematic Diagram of (a) Basic Meta-Block and (b) Compact Meta-Block. The bias term is omitted for clarity.

Model

CIFAR10

CIFAR100

Baseline

Baseline + MCL

Baseline + Aug

Baseline + Aug + MCL

Baseline

Baseline + MCL

Baseline + Aug

Baseline + Aug + MCL

k{=}1

k{=}3

k{=}3(b)

k{=}5(b)

k{=}1

k{=}3

k{=}3(b)

k{=}5(b)

ResNet20

91.43

93.52\pm 0.07

92.33

94.53\pm 0.18

\mathbf{95.32\pm 0.08}

94.66\pm 0.09

95.20\pm 0.05

69.95

70.41\pm 0.05

68.89

72.83\pm 0.19

77.97\pm 0.23

77.32\pm 0.27

\mathbf{78.58\pm 0.23}

ResNet56

93.03

94.32\pm 0.08

94.31

95.81\pm 0.13

96.09\pm 0.12

95.77\pm 0.04

\mathbf{96.12\pm 0.08}

71.97

73.54\pm 0.24

75.26

78.26\pm 0.16

80.37\pm 0.08

79.66\pm 0.18

\mathbf{81.14\pm 0.13}

ResNet110

93.57

94.72\pm 0.18

94.90

96.16\pm 0.00

\mathbf{96.57\pm 0.06}

96.32\pm 0.07

96.43\pm 0.15

74.80

75.00\pm 0.18

78.11

79.73\pm 0.14

81.81\pm 0.19

80.93\pm 0.11

\mathbf{82.25\pm 0.05}

WideResNet16-10

95.44

96.01\pm 0.04

96.80

97.71\pm 0.00

97.78\pm 0.04

\mathbf{97.80\pm 0.07}

97.65\pm 0.03

78.41

79.96\pm 0.26

82.86

84.62\pm 0.08

85.10\pm 0.25

\mathbf{85.62\pm 0.11}

85.44\pm 0.04

TABLE I: Classification accuracy (%) on CIFAR-10 and CIFAR-100 using ResNet and WideResNet backbones with and without the proposed MCL. We report Top-1 accuracy

\pm

STD of 3 independent runs for different meta-kernel sizes

k

and training settings.

Fig. 3: This figure illustrates the comparison between accuracy vs parameters (M) for varying meta-kernel sizes of MCL on the CIFAR-100 dataset.

Models	Params	Top-1 (%)	Top-5 (%)
ResNet18	11.69M	70.25	89.38
+ SE	11.78M	70.98 ( $\uparrow$ 0.73)	90.03 ( $\uparrow$ 0.65)
+ CBAM	11.78M	71.01 ( $\uparrow$ 0.76)	89.85 ( $\uparrow$ 0.47)
+ ECA	11.69M	70.60 ( $\uparrow$ 0.35)	89.68 ( $\uparrow$ 0.30)
+ CGC	11.69M	71.60 ( $\uparrow$ 1.35)	90.35 ( $\uparrow$ 0.97)
+ WeightNet	11.93M	71.56 ( $\uparrow$ 1.31)	90.38 ( $\uparrow$ 1.00)

TABLE II: Performance comparison on the ImageNet dataset using ResNet18 backbone.

Model	ResNet50			ResNet101
	Params	Top-1 (%)	Top-5 (%)	Params	Top-1 (%)	Top-5 (%)
Baseline	25.56M	76.23	93.01	44.55M	77.41	93.67
+ SE	28.07M	77.31 ( $\uparrow$ 1.08)	93.63 ( $\uparrow$ 0.62)	49.29M	78.42 ( $\uparrow$ 1.01)	94.15 ( $\uparrow$ 0.48)
+ CBAM	28.07M	77.46 ( $\uparrow$ 1.23)	93.59 ( $\uparrow$ 0.58)	49.30M	78.50 ( $\uparrow$ 1.09)	94.20 ( $\uparrow$ 0.53)
+ ECA	25.56M	77.34 ( $\uparrow$ 1.11)	93.64 ( $\uparrow$ 0.63)	44.55M	78.60 ( $\uparrow$ 1.19)	94.34 ( $\uparrow$ 0.67)
+ ODConv	90.67M	78.52 ( $\uparrow$ 2.29)	94.01 ( $\uparrow$ 1.00)	90.44M	79.27 ( $\uparrow$ 1.86)	94.47 ( $\uparrow$ 0.80)

TABLE III: Performance comparison on ImageNet dataset using ResNet50 and ResNet101 backbone network.

Dataset	Class	Train	Test
NABirds [ 40 ]	$555$	$23929$	$24633$
CUB-200-2011 [ 41 ]	$200$	$5994$	$5794$
Oxford 102 Flowers [ 42 ]	$102$	$2040$	$6179$
Oxford-IIIT Pet [ 43 ]	$37$	$3680$	$3669$

TABLE IV: Details of the FGVC datasets used in our experiments.

Dataset	Method	Backbone	Accuracy (%)
NABirds	IELT	ViT-B/16	$90.8$
	MP-FGVC	ViT-B/16	$91.0$
	ACC-ViT	ViT-B/16	$91.4$
	MPSA	Swin-B	$92.5$
	MCL $(k=1)$	Swin-B	$\mathbf{92.8}$
CUB-200-2011	MP-FGVC	ViT-B/16	$91.8$

TABLE V: Performance comparison of the proposed MCL on FGVC task using commonly used FGVC datasets.

Backbone	Method	CUB	NABirds	Flowers	Pet
Swin-B	Baseline	$91.93$	$91.99$	$99.41$	$95.09$
	MCL $(k=1)$	$92.72$	$92.83$	$99.64$	$95.91$
	Improvement	$0.79$	$0.84$	$0.24$	$0.82$
ViT-B/16	Baseline	$88.90$	$87.55$	$97.06$	$94.44$
	MCL $(k=1)$	$90.35$	$89.38$	$99.17$	$94.55$
	Improvement	$1.45$	$1.83$	$2.11$	$0.11$

TABLE VI: Performance of the proposed MCL across various FGVC benchmarks using different transformer-based architectures.

Fig. 4: Visual comparison of the proposed method using Grad-Cam visualization.

Method	Backbone	Input	Param	FLOPs
Method	Backbone	Size	(M)	(G)
ViT	ViT-B/16	$448$	$86.4$	$78.5$
IELT	ViT-B/16	$448$	$93.5$	$73.2$
TransFG	ViT-B/16	$448$	$86.4$	$130.2$
RAMS-Trans	ViT-B/16	$448$	$86.4$	$157.4$
ACC-ViT	ViT-B/16	$448$	$87.0$	$162.9$

TABLE VII: Comparison of computational cost of the proposed method while using transformer architectures.

	$k=1$		$k=3$		$k=5$
	#param	Top-1	#param	Top-1	#param	Top-1
$n=1$	$0.35$ M	$71.94$	$3.63$ M	$77.69$	$24.35$ M	$77.41$
$n=2$	$0.42$ M	$72.86$	$3.70$ M	$77.88$	$24.42$ M	$77.92$
$n=3$	$0.49$ M	$\mathbf{73.66}$	$3.78$ M	$\mathbf{78.19}$	$24.49$ M	$\mathbf{77.94}$

TABLE VIII: Classification accuracy of ResNet-20 backbone network using MCL with varying meta-kernel size

k

and order of polynomial expansion

2^{n}

on the CIFAR-100 dataset. Here, MCL is constructed with only the basic meta-block.

	$k=1$		$k=3$		$k=5$
	#param	Top-1	#param	Top-1	#param	Top-1
$n=1$	$0.28$ M	$\mathbf{70.93}$	$0.59$ M	$76.13$	$2.49$ M	$77.83$
$n=2$	$0.28$ M	$70.38$	$0.60$ M	$\mathbf{76.57}$	$2.49$ M	$\mathbf{78.31}$
$n=3$	$0.29$ M	$70.64$	$0.60$ M	$75.86$	$2.50$ M	$78.07$

TABLE IX: Classification accuracy of ResNet-20 backbone network using MCL with varying meta-kernel size

k

and order of polynomial expansion

2^{n}

on the CIFAR-100 dataset. Here, MCL is constructed with only the compact meta-block.

Model	Latency (ms) $\downarrow$	Throughput (images/sec) $\uparrow$	Peak Mem (MB) $\downarrow$
ResNet18	$5.23\pm 1.56$	$24472.18$	$910.82$
+ KW $(4\times)$	$91.67\pm 0.56$	$1396.30$	$2736.41$
+ ODConv $(4\times)$	$35.04\pm 1.71$	$3653.18$	$6376.81$
+ MCL $(k=1)$	$7.68\pm 1.86$	$16659.78$	$992.38$
+ MCL $(k=3)$	$6.75\pm 2.13$	$18957.37$	$1909.15$
ResNet50	$10.24\pm 1.82$	$12503.45$	$1622.69$

TABLE X: Latency and memory overhead comparison of the proposed method on ImageNet test set.

\downarrow

indicates lower is better and

\uparrow

indicates higher is better.

Fig. 5: Convergence behavior of MCL on ImageNet and CIFAR-100 under different meta-kernel sizes. The first two panels show the test-error trajectories of ResNet18, ResNet50, and ResNet101 equipped with MCL on ImageNet for $k=1$ and $k=3$ , respectively. The last two panels show the corresponding CIFAR-100 results for ResNet20, ResNet56, ResNet110, WRN16-10, and WRN28-10. For CIFAR-100, solid lines denote the mean of three independent runs and the shaded bands indicate the standard deviation. Across datasets, architectures, and kernel settings, MCL exhibits smooth convergence and low run-to-run variability, supporting the claim that the proposed layer is stable and easy to optimize in practice.

Explore similar work

Activation-Free Backbones for Image Recognition: Polynomial Alternatives within MetaFormer-Style Vision Models
May 20, 2026Jeffrey Wang, Jonathan Gregory, Grigorios G. ChrysosPolynomial Neural Networks Neural Network Activation Functions
University of Wisconsin–Madison, Madison, WI, USA.
Dynamic Short Convolutions Improve Transformers
Jun 2, 2026Oliver Sieberling, Bharat Runwal, Rameswar Panda +1Dynamic Convolution Transformer
Massachusetts Institute of Technology · MIT-IBM Watson AI Lab
Muon Is Theoretically Wrong For Convolutions, But Empirically Effective
Oct 5, 2026Thibaut Boissin, Thomas Massena, Mathieu Serrurier +1Newton-Schulz Iteration Neural Network Optimization
IRIT · IRIT, DTIPG - SNCF, UT3 · Institut de Recherche Technologique Saint-Exupery, France

May 20, 2026, cs.CVJ/K move · Enter open · S save

Activation-Free Backbones for Image Recognition: Polynomial Alternatives within MetaFormer-Style Vision Models

Jeffrey Wang, Jonathan Gregory, Grigorios G. Chrysos

University of Wisconsin–Madison, Madison, WI, USA.

Polynomial Neural Networks Neural Network Activation Functions Image Classification

Open paper AskPDF arXiv

Modern vision backbones treat pointwise activations (e.g., ReLU, GELU) and exponential softmax as essential sources of nonlinearity, but we demonstrate they are not required within MetaFormer-style vision backbones. We design activation-free polynomial alternatives for three core primitives (MLPs, convolutions, and attention), where Hadamard products replace standard nonlinearities to yield polynomial functions of the input. These modules integrate seamlessly into existing architectures: instantiated within MetaFormer, a modular framework for vision backbones, our PolyNeXt models match or exceed activation-based counterparts across model scales on ImageNet classification, ADE20K semantic segmentation, and out-of-distribution robustness. We also substantially outperform prior polynomial networks at reduced computational cost, showing that polynomial variants of standard modules beat complex custom architectures.

Abstract

Explore similar work

Activation-Free Backbones for Image Recognition: Polynomial Alternatives within MetaFormer-Style Vision Models

Dynamic Short Convolutions Improve Transformers

Muon Is Theoretically Wrong For Convolutions, But Empirically Effective