Paper Reading: Deep Residual Learning for Image Recognition

ResNet
Deep Residual Learning for Image Recognition

Section 1 & 2

1. Depth of a network is of crucial importance, but vanishing/exploding gradients becomes a great obstacle .
2. In some cases, deepen the network layers causes worse performance.
3. Their extremely deep residual nets are easy to optimize, and can easily enjoy accuracy gains from greatly increased depth.

Section 3

It’s sorry that this part doesn’s have many stories to tell.
They found that this architecture with a forward pipe is easier to train, especially when network gets deeper.
Thus, they build a thinner VGG net with deeper layers and get a much better result.

Section 4

Here comes a Bottleneck architecture:

Then there are some comparasions between other networks.Not to mention, ResNet performs great.
Actually, after this paper, we comes to realize the importance of identity mapping. And the method bottleneck architecture is used by various models.