Spatially Separable Convolution
Description
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
Papers Using This Method
SepHRNet: Generating High-Resolution Crop Maps from Remote Sensing imagery using HRNet with Separable Convolution2023-07-11EEEA-Net: An Early Exit Evolutionary Neural Architecture Search2021-08-13Neural Audio Fingerprint for High-specific Audio Retrieval based on Contrastive Learning2020-10-22RelativeNAS: Relative Neural Architecture Search via Slow-Fast Learning2020-09-14GDP: Generalized Device Placement for Dataflow Graphs2019-09-28Learning Data Augmentation Strategies for Object Detection2019-06-26HAWQ: Hessian AWare Quantization of Neural Networks with Mixed-Precision2019-04-29NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection2019-04-16ANODE: Unconditionally Accurate Memory-Efficient Gradients for Neural ODEs2019-02-27GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism2018-11-16SqueezeNext: Hardware-Aware Neural Network Design2018-03-23Regularized Evolution for Image Classifier Architecture Search2018-02-05