Research map / Architectures and efficiency / CNN design

KernelDNA: Cross-Layer Kernel Sharing via Decoupled Neural Adapters

NeurIPS 2026 · Architectures and efficiency › CNN design · approach: Large and dynamic kernel design

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes KernelDNA, a dynamic convolution variant that shares kernels across layers via decoupled neural adapters to reduce parameters without sacrificing inference throughput.

Problem

Dynamic convolution adds representational flexibility to CNNs but either inflates parameters linearly with the number of base kernels or reduces inference throughput due to runtime kernel assembly.

Approach

Exploits cross-layer redundancy: CKA analysis shows convolutional kernels within a stage are highly similar, and uses cross-layer kernel sharing through decoupled neural adapters instead of per-layer base kernels with runtime assembly.

Surprising findings

• Within-stage convolutional kernels across modern CNN families are highly similar per CKA analysis, indicating substantial cross-layer redundancy.

Experimental setup

Baselines: dynamic convolution

Builds on: dynamic convolution

Closest papers by meaning

• KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.91)
• Approximation and Learning with Deep Convolutional Models: a Kernel Perspective (ICLR 2022 · similarity 0.90)
• End-to-End Kernel Learning with Supervised Convolutional Kernel Networks (NeurIPS 2016 · similarity 0.90)
• Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.89)
• PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · similarity 0.87)
• Unveiling the Unseen: Identifiable Clusters in Trained Depthwise Convolutional Kernels (ICLR 2024 · similarity 0.86)