16. MnasNet: Platform-Aware Neural Architecture Search for Mobile
• Mingxing Tan et al., Google Brain
• arXiv:1807.11626v1
16
紹介する論文
• 性能だけでなく,modelの処理速度も考慮した多目的構造最適化
手法
• mobile phoneでの実行速度を最適化に使用
I will talk about image restoration using evolutionary search.
Image restoration is to recover a clean image from its corrupted version.
These are image restoration tasks, image inpainting and denoising.
In order to solve this task, learning-based methods which use CNNs have been introduced, and have shown good performance.
In these studies, researchers have approached the problem mainly from two directions.
One is to design new network architectures.
For example, the network of the bottom left is called MemNet, which contains many recursive connections and gate units.
The other is to develop new loss functions or training methods.
A recent trend is to use adversarial training, in which a generator is trained to perform image restoration, and a discriminator is trained to distinguish whether an input image is true image or a recovered one.