date | tags |
---|---|
2022-04-16 |
paper, deep-learning, survey, tinyml |
Yan Zhou, Shaochang Chen, Yiming Wang, Wenming Huan
ITOEC 2020
Year: 2020
This paper summarizes some of the most impactful ideas on the field of lightweight convolutional neural networks.
First of all, the paper introduces 4 types of efficient convolutions.
- Dilated convolutions: consisting of injecting holes into the convolution kernel to effectively increasing the receptive field without increasing the number of parameters
- Deformable convolutions: consisting of letting the network choose the shape of the kernel, moving away from the regular rectangular size
- Group convolutions: consisting of distributing the computation of the convolutional layers per group, as a way to distribute the calculation and achieving a sparser computation (all input channels no longer depending on all output channels)
- Depthwise separable convolutions: consisting of unfolding the classical convolutional layer operation into two composed operations: a pointwise convolution capable of modifying the number of channels, and a depthwise convolution applied to analyse spatial information.
Then, the authors take 6 examples of neural networks that leaded the trend of current efficient CNNs:
- Squeezenet: 1x1 convolutions used to reduce the computation
- MobileNetV1: depthwise separable convolutions
- MobileNetV2: depthwise separable convolutions + linear bottleneck
- ShuffleNet: group convolutions with channel mixing operation
- MixNet: depthwise separable convolutions + mixed kernel sizes
- EfficientNet: efficient choice of depth, width and resolution