Approach
The DNN is built on a new basic module Conv-M that provides more diverse feature extractors without significantly increasing parameters. The unified framework of our DA method will simultaneously learn invariance across domains, reduce divergence of feature representations and adapt label prediction. Our DNN has 4.1M parameters—only 6.7% of AlexNet or 59% of GoogLeNet. Experiments show that our DNN obtains GoogLeNet-level accuracy both on classification and DA, and our DA method slightly outperforms previous competitive ones.
The design idea is to capture more diverse details at different levels, while using fewer parameters. To achieve this goal, the dilated convolution for multi-resolution and deconvolution are introduced. The dilated convolution can extract features with a larger receptive field without increasing the kernel size, extracting features from a 5×5 window with a 3×3 kernel. The deconvolution is to reconstruct shapes of the input, providing distinct features from regular convolution.
The detailed design of Conv-M shows that the input feature maps from the previous layer are respectively processed by regular convolution (conv), dilated convolution (dilated conv) and deconvolution (deconv) in three branches. Their outputs will be concatenated together.
Experiment
References：
A Compact DNN: Approaching GoogLeNet-Level Accuracy of Classification and Domain Adaptation， Chunpeng Wu， 2017，CVPR
