# Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved MobileNets

arXiv: https://arxiv.org/abs/2003.13549 · HTML: https://mlsanity.com/paper/2003.13549/

## Key facts

- Cited by 28 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 30 Mar 2020.

## Built on by (most cited first, 28 of 28)

- [Run, Don't Walk: Chasing Higher FLOPS for Faster Neural Networks](https://mlsanity.com/paper/2303.03667/) (2023)
- [Multi-Scale High-Resolution Vision Transformer for Semantic Segmentation](https://mlsanity.com/paper/2111.01236/) (2021)
- [Contextual Transformer Networks for Visual Recognition](https://mlsanity.com/paper/2107.12292/) (2021)
- [Blueprint Separable Residual Network for Efficient Image Super-Resolution](https://mlsanity.com/paper/2205.05996/) (2022)
- [The Tenth NTIRE 2025 Efficient Super-Resolution Challenge Report](https://mlsanity.com/paper/2504.10686/) (2025)
- [Swift Parameter-free Attention Network for Efficient Super-Resolution](https://mlsanity.com/paper/2311.12770/) (2023)
- [Exploiting Redundancy: Separable Group Convolutional Networks on Lie Groups](https://mlsanity.com/paper/2110.13059/) (2021)
- [NeurOLight: A Physics-Agnostic Neural Operator Enabling Parametric Photonic Device Simulation](https://mlsanity.com/paper/2209.10098/) (2022)
- [A Comprehensive Survey of Convolutions in Deep Learning: Applications, Challenges, and Future Trends](https://mlsanity.com/paper/2402.15490/) (2024)
- [Large Kernel Distillation Network for Efficient Single Image Super-Resolution](https://mlsanity.com/paper/2407.14340/) (2024)
- [Oriented object detection in optical remote sensing images using deep learning: a survey](https://mlsanity.com/paper/2302.10473/) (2023)
- [Exploiting Learned Symmetries in Group Equivariant Convolutions](https://mlsanity.com/paper/2106.04914/) (2021)
- [Towards Memory-Efficient Neural Networks via Multi-Level in situ Generation](https://mlsanity.com/paper/2108.11430/) (2021)
- [Convolutional Networks with Oriented 1D Kernels](https://mlsanity.com/paper/2309.15812/) (2023)
- [Foundation Model for Lossy Compression of Spatiotemporal Scientific Data](https://mlsanity.com/paper/2412.17184/) (2024)
- [Quality Assessment of Low Light Restored Images: A Subjective Study and an Unsupervised Model](https://mlsanity.com/paper/2202.02277/) (2022)
- [Large coordinate kernel attention network for lightweight image super-resolution](https://mlsanity.com/paper/2405.09353/) (2024)
- [Implicit Feature Decoupling with Depthwise Quantization](https://mlsanity.com/paper/2203.08080/) (2022)
- [U(1) Symmetry-breaking Observed in Generic CNN Bottleneck Layers](https://mlsanity.com/paper/2206.02220/) (2022)
- [Tech Report: One-stage Lightweight Object Detectors](https://mlsanity.com/paper/2210.17151/) (2022)
- [LIB-KD: Teaching Inductive Bias for Efficient Vision Transformer Distillation and Compression](https://mlsanity.com/paper/2310.00369/) (2023)
- [Coloring Deep CNN Layers with Activation Hue Loss](https://mlsanity.com/paper/2310.03911/) (2023)
- [Deep Space Separable Distillation for Lightweight Acoustic Scene Classification](https://mlsanity.com/paper/2405.03567/) (2024)
- [MOGNET: A Mux-residual quantized Network leveraging Online-Generated weights](https://mlsanity.com/paper/2501.09531/) (2025)
- [Involution and BSConv Multi-Depth Distillation Network for Lightweight Image Super-Resolution](https://mlsanity.com/paper/2503.14779/) (2025)
- [Dark Channel-Assisted Depth-from-Defocus from a Single Image](https://mlsanity.com/paper/2506.06643/) (2025)
- [LKFMixer: Exploring Large Kernel Feature For Efficient Image Super-Resolution](https://mlsanity.com/paper/2508.11391/) (2025)
- [DPFNAS: Differential Privacy-Enhanced Federated Neural Architecture Search for 6G Edge Intelligence](https://mlsanity.com/paper/2509.23030/) (2025)

## Abstract (arXiv)

We introduce blueprint separable convolutions (BSConv) as highly efficient building blocks for CNNs. They are motivated by quantitative analyses of kernel properties from trained models, which show the dominance of correlations along the depth axis. Based on our findings, we formulate a theoretical foundation from which we derive efficient implementations using only standard layers. Moreover, our approach provides a thorough theoretical derivation, interpretation, and justification for the application of depthwise separable convolutions (DSCs) in general, which have become the basis of many modern network architectures. Ultimately, we reveal that DSC-based architectures such as MobileNets implicitly rely on cross-kernel correlations, while our BSConv formulation is based on intra-kernel correlations and thus allows for a more efficient separation of regular convolutions. Extensive experiments on large-scale and fine-grained classification datasets show that BSConvs clearly and consistently improve MobileNets and other DSC-based architectures without introducing any further complexity. For fine-grained datasets, we achieve an improvement of up to 13.7 percentage points. In addition, if used as drop-in replacement for standard architectures such as ResNets, BSConv variants also outperform their vanilla counterparts by up to 9.5 percentage points on ImageNet. Code and models are available under https://github.com/zeiss-microscopy/BSConv.