卷积神经网络Quiz2

Question 1
Which of the following do you typically see as you move to deeper layers in a ConvNet?

nH and nW increases, while nC also increases

nH and nW increases, while nC decreases

nH and nW decrease, while nC increases

nH and nW decreases, while nC also decreases
解析：经典模式：随着网络的深度增加，图像的大小在缩小，与此同时，通道的数量却在增加；

---

Question 2
Which of the following do you typically see in a ConvNet? (Check all that apply.)

Multiple CONV layers followed by a POOL layer

Multiple POOL layers followed by a CONV layer

FC layers in the last few layers

FC layers in the first few layers
解析：比如在经典的AlexNet网络中，就存在多个卷积层后跟着一个池化层；几乎所有的网络结构都是在最后几层使用全连接；第二个选项不清楚，个人觉得不对，因为“最大池化是对前一层得到的特征图进行池化减小，仅由当前小区域内的最大值来代表最终池化后的值。”，使用多个池化和使用一个较大的池化，效果应该一样

---

Question 3
In order to be able to build very deep networks, we usually only use pooling layers to downsize the height/width of the activation volumes while convolutions are used with “valid” padding. Otherwise, we would downsize the input of the model too quickly.

True

False

---

Question 4
Training a deeper network (for example, adding additional layers to the network) allows the network to fit more complex functions and thus almost always results in lower training error. For this question, assume we’re referring to “plain” networks.

True

False
解析：训练层数的增加并不一定会带来更小的误差

在没有残差的普通神经网络中，训练的误差实际上是随着网络层数的加深，先减小再增加；
在有残差的ResNet中，即使网络再深，训练误差都会随着网络层数的加深逐渐减小。

---

Question 5
The following equation captures the computation in a ResNet block. What goes into the two blanks above?

0 and z^[l+1], respectively

a^[l] and 0, respectively

0 and a^[l], respectively

z^[l] and a^[l], respectively

---

Question 6
Which ones of the following statements on Residual Networks are true? (Check all that apply.)

The skip-connections compute a complex non-linear function of the input to pass to a deeper layer in the network.

Using a skip-connection helps the gradient to backpropagate and thus helps you to train deeper networks

A ResNet with L layers would have on the order of L² skip connections in total.

The skip-connection makes it easy for the network to learn an identity mapping between the input and the output within the ResNet block.
解析：

可以看出，skip-connection可以计算更加复杂的非线性函数，同时，当压缩W和b的值时，也更容易实现恒等式

---

Question 7
Suppose you have an input volume of dimension 64x64x16. How many parameters would a single 1x1 convolutional filter have (including the bias)?

2

17

4097

1
解析：

1×1卷积层计算成本：28×28×16×1×1×192=2.4M
5×5卷积层计算成本：28×28×32×5×5×16=10.0M
在本例中，为64*64*1*1*1*1+1=4097

---

Question 8
Suppose you have an input volume of dimension nH x nW x nC. Which of the following statements you agree with? (Assume that “1x1 convolutional layer” below always uses a stride of 1 and no padding.)

You can use a pooling layer to reduce nH, nW, and nC.

You can use a pooling layer to reduce nH, nW, but not nC.

You can use a 1x1 convolutional layer to reduce nC but not nH, nW.

You can use a 1x1 convolutional layer to reduce nH, nW, and nC.
解析：pooling layer只减少nH, nW，1x1 convolutional layer则让nH, nW保持不变，减下或不变nC

---

Question 9
Which ones of the following statements on Inception Networks are true? (Check all that apply.)

Inception blocks usually use 1x1 convolutions to reduce the input data volume’s size before applying 3x3 and 5x5 convolutions.

A single inception block allows the network to use a combination of 1x1, 3x3, 5x5 convolutions and pooling.

Inception networks incorporates a variety of network architectures (similar to dropout, which randomly chooses a network architecture on each step) and thus has a similar regularizing effect as dropout.

Making an inception network deeper (by stacking more inception blocks together) should not hurt training set performance.

---

Question 10
Which of the following are common reasons for using open-source implementations of ConvNets (both the model and/or weights)? Check all that apply.

The same techniques for winning computer vision competitions, such as using multiple crops at test time, are widely used in practical deployments (or production system deployments) of ConvNets.

Parameters trained for one computer vision task are often useful as pretraining for other computer vision tasks.

It is a convenient way to get working an implementation of a complex ConvNet architecture.

A model trained for one computer vision task can usually be used to perform data augmentation even for a different computer vision task.
解析：为什么采用开源实现，1是因为可以更方便的实现该网络（自己实现可能较为困难），2是因为有些时候有些权重参数可以直接拿来使用

---

参考：http://blog.csdn.net/koala_tree/article/details/78531398