Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
This paper discusses the problems with training very deep networks and introduces “residual connections” to alleviate the problem. The authors assert that a deep net with layers copied from a shallow network, and the remaining layers all assigned identity function should be no worse than the shallow network. In practice however, optimizers are unable to find even this solution (let alone a better one), and actually degrade in performance.
The intermediate layers are made to learn “residual functions” ie. H(x)-x, where x is the input to the layers and H(x) is the layer we actually want to learn. The reasoning is that the outputs are equivalent but the residual connections are easier to learn.
The authors also test their Resnets for object detection by bolting it on to the Faster-RCNN framework and show SOTA results.
Things I learned
- Degradation problem
- Increasing depth beyond a certain point degrades both training and test error, thus highlighting that overfitting is not the problem (one would expect at least training error to be less than that of the shallow network)
- The authors assert that identity mappings are hard to learn for a stack of nonlinear layers
- They also found through experiments that this is not due to the vanishing gradient problem (They use BN and verify that forward/backward signals don’t vanish).
- More analysis in Identity Mappings in Deep Residual Networks
- 11 convolution layers are used for dimensionality reduction- spatial resolution maintained, but the number of feature maps can be reduced (volume) by using fewer 11 conv filters than the current volume. Introduced in the Inception architecture for reducing feature map size before blowing them up again with 33 or 55 conv layers.
- How are the shortcut connections actually implemented in a software package?
- x-> intermediate layer-> F(x)+x. How is the intermediate layer forced to produce just F(x)? For instance, is it frozen
- In practice, why are Resnets better than Highway Networks (Srivastava et al.) when the identity mapping is a subset of the solution space of the Highway networks?
- It doesn’t have to be forced or anything. It can be seen as an additional layer, where the output of the intermediate layers, F(x) is summed with x
- Unclear. Maybe it’s something about learnability of identity mappings