2019 AAAI Transferable Attention for Domain Adaptation

没写完

2.Related Works

2.1 Adversarial Learning

2.2 Global and Local Attention

2.3 Transferable Attention for Domain Adaptation

Definition 1 (Transferable Attention)

可迁移attention（TA） focus on ： similarity or distinction between
two objects—to reason about whether they are transferable.
提出重要的事实 the attention may be multi-scale,多尺度的
local 对应regions（a specific part of an object.），global对应the whole

Transferable Local Attention

local discriminator

回顾了DANN
local attention 突出了部分region，相对于background。
怎么实现的呢？
把$G_d$Gd​分成$K$K个region-wise 的$G_d^k$Gdk​。
怎么分的K个region的呢？

从Fig1看，是分成方格，
K由network architecture决定,
求和再平均，$n=n_s+n_t$n=ns​+nt​是所有样本的个数，$K=W\times H$K=W×H是最后一层卷积层feature map的w和h。
举个例子ResNet-50: $7 \times 7 \times 2048$7×7×2048,$K=49$K=49

用熵来度量可迁移性

local attention value for each region k

为了robust，参考2017 CVPR Residual attention network for image classification加了residual connection，
the locally attended features $h^k_i$hik​

Transferable Global Attention

global discriminator