Notes for Deep Learning: Optimization Algorithms
This post is an overview of different optimization algorithms for neural networks.
Notes for Object Detection: One Stage Methods
In this post, we focus on two mainstreams of one-stage object detection methods: YOLO family and SSD family. Compared to two-stage methods (like R-CNN series), those models skip the region proposal stage and directly extract detection results from feature maps. For that reason, one-stage models are faster but at the cost of reduced accuracy.
An Overview of DCNN Architectures: Efficient Models
In this post, we discuss the computally efficient DCNN architectures, such as MobileNet, ShuffleNet and their variants.
From ProGAN to StyleGAN
In this post, we are looking into two high-resolution image generation models: ProGAN and StyleGAN. They generates the artificial images gradually, starting from a very low resolution and continuing to a high resolution (finally $1024\times 1024$).
Notes for Adversarial Discriminative Domain Adaptation
About this paper
- Title: Adversarial Discriminative Domain Adaptation
- Authors: Eric Tzeng, Judy Hoffman, Kate Saenko, Trevor Darrell
- Topic: Domain Adaptation
- From: arXiv:1702.05464, appearing in CVPR 2017
Contributions
- 将之前的论文里提到的一些方法，例如weight sharing、base models、adversarial loss等，归入了统一的框架之中，并进行了测试；
- 提出了一种新的框架ADDA，主要思想是不做分类器的自适应，而是设法将目标域的数据映射到域源域差不多的特征空间上，这样就能够复用源域的分类器。