Bio-RadDS: Biologically Inspired Radially Balanced Dense-to-Sparse Convolution for Efficient Large-Kernel Vision Models

Large-kernel convolutions capture broad spatial context, but their computational cost increases rapidly with kernel area and many spatial connections are essentially redundant. Existing methods either retain dense spatial execution or rely on predefined sparse patterns, limiting their ability to learn efficient spatial connectivity from data. To this end, we present Bio-inspired Radially Balanced Dense-to- Sparse Convolution (Bio-RadDS), a biologically inspired operator that keeps the kernel center dense and learns sparse connections across peripheral radial bands. We combine learnable spatial modulation with a gradual dense-to-sparse transition to obtain compact connectivity while preserving broad spatial coverage. The resulting fixed topology is paired with dedicated sparse execution to enable practical inference acceleration. In the principal operator experiment, Bio-RadDS removes about 60% of the connections in the target kernels and ranks first in 15 of 21 settings across seven image classification datasets. On ConvNeXt-Tiny, the FP32 sparse deployment achieves 2.5×–3.5× end-to-end acceleration as the input size increases from 224 to 2048. These results show that Bio-RadDS learns compact large-kernel connectivity while maintaining strong classification accuracy and enabling efficient deployment at demanding input resolutions.