Chatting about neural networks-writing to beginners (1)

Preface: Keep your style consistent. Before you officially start writing, start with a long talk. There are too many books and articles about neural networks, so I am not allowed to talk about them in a word that is too arrogant. I try to write a little more information. After reading this article, I can have a general understanding of neural networks and have some experience. Next, let's take a look at the big books and exercises based on our own questions. If you can combine it with your own learning and work problems, it will be excellent.

================== Start writing

I. Overall understanding of Neural Networks

What is a neural network? Before you fall into a variety of concepts, terms, and theories, be sure to remind yourself that only by fully understanding what a neural network is, you will not lose your way in the future.

See the following three figures:

　　　　

Figure 1 Figure 2 Figure 3

These three images provide a better description of the neural network:The neural network is a black box, just like the human brain.

(1) We will provide some input and output information for this black box. This black box will be able to automatically find the relationship between the input and the output. When we enter something later, they can give us some output. Just like the human brain, we constantly input people and output their corresponding gender. When our brains learn and grow up, we can see other people and then identify their gender.

(2) Some inputs may not necessarily have outputs, but can find the relationship between these inputs. Just like the human brain, we look at people and input all kinds of people. As our human brain keeps learning and training, it will automatically divide them into beautiful, common, and ugly relationships.

Therefore, it is quite appropriate to call it a neural network. The following is an analogy:

(1) the human brain is both intelligent and stupid. Therefore, the neural networks we construct in the future are also intelligent and stupid. Only constant training and constant adjustment of appropriate parameters are required, in order to make our neural networks more and more intelligent.

(2) All human brains have things they are suitable for. For example, some people are suitable for science, some are suitable for liberal arts, some are suitable for economics, and some are suitable for politics .. Similarly, a variety of neural network models have their own things that are suitable for you. Only by understanding the basic principles of neural networks can you select a suitable network based on your actual needs. (I suddenly think of a saying a few days ago when I was a child: "after so many years of TMD, I found that I am not suitable for school ").

(3) When there is a false positive in the human brain, neural networks cannot be 100% accurate. Therefore, neural networks cannot be used for anything that requires accuracy, A precise mathematical model and formula are required for calculation.

In a word, "neural networks are a magic black box ". Next, let's open the black box and see what's going on in it.

2. Open the Black Box

　　You may have been stunned when you opened the black box. There are some messy things in it. Don't worry. Let's analyze it a little bit.

The magical creator created the human brain, so some people tried to create neural networks to simulate the human brain. What would you do if you wanted to design a neural network? If you cannot understand it, let's take a look at the design of the predecessors. There are many predecessors here, so that everyone canUnderstanding the principles of Neural Networks, AndPractical OperationI divided my predecessors into two categories:Math and programmer. Mathematicians design the overall idea, while programmers use programs to implement what mathematicians call the function (using Matlab as an example ).

Let's see how our predecessors designed it.

　　We found a lot in this black box.Small round ball, As shown in:

These small balls have a nice name called"Neurons", at first you may think that these small balls are arranged in disorder. In fact, if you look at them carefully, they are arranged in a regular manner and different according to their arrangement rules, the neural network structure is roughly divided into the following types:

(1) Feed-forward type

 

These small globes are arranged one by one. Input indicates the input signal. from the beginning to the end, till the output, there is no signal transmission between the neurons and neurons. The first layer is the input unit, the second layer is called the hidden layer, and the third layer is called the output layer (the input unit is not a neuron, so there is a layer 2 neuron in the figure ).

For example, BP neural networks and sensor neural networks are arranged in this order.

 

(2) Feedback type

 

As the name suggests, the final input signal is not only finished, but also fed back the output signal. Therefore, it is called a Feedback Neural Network.

For example, regression BP networks belong to this type.

(3) layer-based Interconnection

In the preceding structure, neurons in the same layer are not associated, while neurons in the same layer are associated with each other, and neurons in the same layer transmit information to each other.

 

How about it? I really appreciate our predecessors and can think of such a method. These small balls (neurons) are connected to each other and they can do a lot of complicated things. But what exactly are these small balls, next, let's focus on this article. let's dive into the ball and see what is going on.

3. Open the small sphere (neuron)

The structure of each neuron is shown in: (Note that there are various formulas and symbols. Remember that the neural network is a black box with many small ball (neuron) connections, they arrange them layer by layer, with various structural methods)

The above figure shows the internal structure of neurons. Although it is not a ball, you can imagine that it has a round outer shell wrapped in it. Let's take a look at what he has.

(1) input: Inputs represents x1, x2,..., XM represents the input signal.Sigma functions of each neuronAfter a sum, we can see that they are directly added, but each input is multiplied by a coefficient w and then added. These w are called weights or weights, or weight.

(2) Bias: we can see that there is not only the input multiplied by weight, but also a bias. The input is 1 and the weight is W0, therefore, it can be regarded as an input signal with a fixed input of 1 and a weight of W0.In other books, the value of bias is represented by Bi symbols, or something else. It should also be noted that this bias has many names, such as the threshold value, threshold, and offset.

(3)Activation function: After the input and bias are summed by Σ, a function is used to process and output the result. This function is called an activation function, which is also called an activation function and a transfer function, that is, transfer function, also called TF. In short, these names are all the same thing.

A magic neural network is made up of neurons one by one. This black box of magic can do a lot of things.

Then the problem arises. Can we change those things to solve all kinds of problems?

What we can change is nothing more than the following:

(0) input and output formats and quality

(1) weight of each neuron W

(2) offset bias. in a broad sense, this change is also considered to change the weight of W0.

(3) activate a function

(4) neural network layers

(5) number of neurons in each layer

(6) neural network structure

Then the question comes again. How can we change these parameters?

This article is too long. Let's continue with the next article and try to give another example.

 

 

======================================

Recently, I helped my friends get a public number. at 06:30 every morning, I sent a 60-second speech. If you are interested, please take a look.

No.: wailixs)

QR code:

 

Chatting about neural networks-writing to beginners (1)