Methods > Computer Vision > Convolutions

Spatially Separable Convolution


A Spatially Separable Convolution decomposes a convolution into two separate operations. In regular convolution, if we have a 3 x 3 kernel then we directly convolve this with the image. We can divide a 3 x 3 kernel into a 3 x 1 kernel and a 1 x 3 kernel. Then, in spatially separable convolution, we first convolve the 3 x 1 kernel then the 1 x 3 kernel. This requires 6 instead of 9 parameters compared to regular convolution, and so it is more parameter efficient (additionally less matrix multiplications are required).

Image Source: Kunlun Bai

Latest Papers

PAPER DATE
RelativeNAS: Relative Neural Architecture Search via Slow-Fast Learning
| Hao TanRan ChengShihua HuangCheng HeChangxiao QiuFan YangPing Luo
2020-09-14
GDP: Generalized Device Placement for Dataflow Graphs
Yanqi ZhouSudip RoyAmirali AbdolrashidiDaniel WongPeter C. MaQiumin XuMing ZhongHanxiao LiuAnna GoldieAzalia MirhoseiniJames Laudon
2019-09-28
Learning Data Augmentation Strategies for Object Detection
| Barret ZophEkin D. CubukGolnaz GhiasiTsung-Yi LinJonathon ShlensQuoc V. Le
2019-06-26
HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision
Zhen DongZhewei YaoAmir GholamiMichael MahoneyKurt Keutzer
2019-04-29
NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection
| Golnaz GhiasiTsung-Yi LinRuoming PangQuoc V. Le
2019-04-16
ANODE: Unconditionally Accurate Memory-Efficient Gradients for Neural ODEs
| Amir GholamiKurt KeutzerGeorge Biros
2019-02-27
GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism
| Yanping HuangYoulong ChengAnkur BapnaOrhan FiratMia Xu ChenDehao ChenHyoukJoong LeeJiquan NgiamQuoc V. LeYonghui WuZhifeng Chen
2018-11-16
SqueezeNext: Hardware-Aware Neural Network Design
| Amir GholamiKiseok KwonBichen WuZizheng TaiXiangyu YuePeter JinSicheng ZhaoKurt Keutzer
2018-03-23
Regularized Evolution for Image Classifier Architecture Search
| Esteban RealAlok AggarwalYanping HuangQuoc V. Le
2018-02-05

Components

COMPONENT TYPE
🤖 No Components Found You can add them if they exist; e.g. Mask R-CNN uses RoIAlign

Categories