intuition

## 2️⃣[Attentive Pooling Networks]

G所代表的意义是Q和A的每个词之间的对齐：对于第i行来说，代表Q的第i个词和A中所有词的一个分数；对于第j列来说，代表第j个词和Q中所有词的分数。

## 3️⃣[Improved Regularization of Convolutional Neural Networks with Cutout]

dropout in convolutional layers simply acts to increase robustness to noisy inputs, rather than having the same model averaging effect that is observed in fully-connected layers

In this sense, cutout is much closer to data augmentation than dropout, as it is not creating noise, but instead generating images that appear novel to the network