English 中文(简体)
利用CNN提取的特征数量
原标题:Number of features extracted using the CNN

最近,我在一次与权力下放网络(CNNs)有关的面谈中遇到了一个问题,我希望澄清我的理解的一个具体方面。

In the context of the following CNN architecture:

    i = Input(shape=x_train[0].shape)
    x = Conv2D(32, (3, 3), strides=2, activation= relu )(i)
    x = Conv2D(64, (3, 3), strides=2, activation= relu )(x)
    x = Conv2D(128, (3, 3), strides=2, activation= relu )(x)
    x = Flatten()(x)
    x = Dropout(0.5)(x)
    x = Dense(1024, activation= relu )(x)
    x = Dropout(0.2)(x)
    x = Dense(K, activation= softmax )(x)
 
    model = Model(i, x)

我正在要求澄清对聚合层中使用的过滤器数量(32、64、128)的解释。 具体来说,我谨确认,这些过滤器的总和是从图像中提取的特征地图总数。 此外,该专题地图的总数是否与从网络获得的特征(低级/高级特征)总数相符? 换言之,能否将陶器视为等同于从图像中提取的特征数目? 可以说,我们在这个构架中提取了224个特征?

我已经走过不同的博客职位,但并没有明确地得到这样的答案,即这一地雷的持久性危险是正确的。

问题回答

我将努力根据我从我的课程中了解到的情况来回答。 我不敢肯定,但也许会帮助你。 在这里,对于有线电视新闻网、方言或过滤器,对我来说是一样的。

在一个有色人种的电离层中,eachfil/kernel设计成在投入数据中添加一个特定类型的特征。 例如,如果在一层有32个过滤器,则每个过滤器都独立扫描输入图像(或前层的产出),以根据网络在培训期间所学到的内容,探测32个不同的特征(如:边缘、文字、肤色或其他图像形态)。 每一过滤器/舱载体的射线是个特征图,这是输入数据的一种表述,突出显示过滤器/油箱用于探测的特征。

在你描述的结构中,三层共聚层分别有32、64和128个过滤器。 必须指出,由convolutional levels制作的特质地图数量等于该层过滤器/strong>。 因此:

The first Conv2D layer with 32 filters produces 32 feature maps.
The second Conv2D layer with 64 filters produces 64 feature maps.
The third Conv2D layer with 128 filters produces 128 feature maps.

这些专题地图是,不是以你可能想的方式在不同层次上添加。 该网络所提取的特征总数并非仅仅是所有层次的过滤器的总和(如果是的话,32 + 64 + 128 = 224)。 相反,each(hidden)层将从前层收到的一套特征地图转化为一套新的特征地图

The output of these filters is a corresponding number of feature maps, each map being the result of one filter applied to the input.

第一层可能会发现边缘或肤色等简单特征(基本模式),而深层层则将这些特征与文字或具体模式等更为复杂的特征(更为抽象)。 特征的深度和复杂性随着各个层次的形成而增加,但特征的数量并不是过滤器的直截了当。

Shortly, the number of kernels in each layer determines the number of feature maps, and hence the number of different features detected at that layer. The total number of features extracted by the network is a more complex concept. It involves the hierarchical and compositional nature of how features are combined and transformed across the layers of the network.





相关问题
Backpropagation issues

I have a couple of questions about how to code the backpropagation algorithm of neural networks: The topology of my networks is an input layer, hidden layer and output layer. Both the hidden layer ...

How to determine for which value artificial neuron will fire?

I m trying to determine for the articial neuron shown below the values (0 or 1) for the inputs i1, i2, and i3 for which it will fire (i0 is the input for the bias weight and will always be -1). The ...

Prototyping neural networks

from your experience, which is the most effective approach to implement artificial neural networks prototypes? It is a lot of hype about R (free, but I didn t work with it) or Matlab (not free), ...

AIML pattern matching - howto?

I m having a problem trying to understand how does AIML pattern matching works. What s the difference between _ and *? And how I should use them to get the best match? I have this document only, ...

Continuous output in Neural Networks

How can I set Neural Networks so they accept and output a continuous range of values instead of a discrete ones? From what I recall from doing a Neural Network class a couple of years ago, the ...

Competitive Learning in Neural Networks

I am playing with some neural network simulations. I d like to get two neural networks sharing the input and output nodes (with other nodes being distinct and part of two different routes) to compete. ...