[introduction]So far, what we pay close attention to is monolayer perception implement, it inputs layer and composition of an output layer by. You still may remember, we use term ” monolayer ” because this configuration includes node of activity of a computation only,be, apply activation function to revise the node of data next through sue for peace namely. The node in inputting a layer just issues data.

So far, what we pay close attention to is monolayer perception implement, it inputs layer and composition of an output layer by. You still may remember, we use term ” monolayer ” because this configuration includes node of activity of a computation only,be, apply activation function to revise the node of data next through sue for peace namely. The node in inputting a layer just issues data.

Monolayer is felt implement very simple on the concept, training course is very simple. Unfortunate is, it does not offer us complex, the function that the applied process place in real life needs. My impression is, explanatory monolayer is felt implement the standard method that restricts basically is to use Boer operation to regard a specification as sexual give typical examples, this is the method that my general uses in the article.

Door of nerve network logic

We will use an extremely complex microprocessor to realize network of a nerve, this nerve network can accomplish as same as the circuit that comprises by a few transistor thing, this idea is a little humorous. But meanwhile, pondered over this problem to emphasize monolayer feeling with this kind of means implement the inadequacy that approachs a tool as general classification and function — if our perception implement cannot duplicate the behavior of individual and logistic door, we know we need to find a better perception implement.

Let us return this series piece the systematic configuration of article intermediary carry on.

This perception implement average figure let me remember a logistic door, in fact, it can become very quickly such. Assume we are used contain input vector element 0 train this network originally with the appearance of 1, and be inputted when two only be equal to 1 when output value just is equal to 1. The result will be network of a nerve, it inputs vector to undertake classified with be similar to the means of the electric behavior with the door to be opposite.

The dimension that this network inputs is 2, accordingly we will input example scale easily to become 2 dimension figure. Hypothesis Input0 is corresponding at horizontal axis, input1 is corresponding at vertical axis. 4 kinds of possible input combination will arrange as follows:

Because we are being operated in duplicate AND, because this network needs to revise its weight, make input vector [1, the output of 1] is 1, other 3 output that input vector are 0. Be based on these information, let us will input a space to differentiate to need to export classified part at place for correspondence:

Linear can distribute data

Institute of as above plan is shown, when we realize AND operation, can adopt scale one undertakes classified to the input vector of scale point-blank. The output value that all content of line a side receive is 1, and the output value that all content of another side receive is 0. Accordingly, below the circumstance of AND operation, the data that presents a network is linear disjunctive. OR operation also is such:

The fact proves, monolayer is felt implement in data linear only disjunctive circumstance issues ability to solve a problem. No matter input the dimension of example how, it is such. 2 dimension circumstance is very easy visible, because we are OK,scale is nodded and part them with a line. To generalize the concept of linear separability, we must be used ” exceed plane ” this word is not ” line ” . Exceeding plane is OK in N dimension space the geometrical feature of separate data. In 2 dimension environment, exceeding plane is one dimension feature (namely a line) . In three-dimensional environment, exceeding plane is a common 2 dimension plane. In N dimension environment, exceed plane to have (N -1) dimension.

Solve cannot the problem of linear depart

In training a process, monolayer is felt implement the position that use training appearance determines to classification exceeds plane originally. After finding what can distribute data reliably into correct classified category to exceed plane, it can take action. However, if exceed plane nonexistent, perception implement will not find it. Let us see the input of linear depart outputs an example that cannot concern:

Do you admit this kind of relation? Have a look again, no more than is different or operation. You cannot use data of linear space XOR. Accordingly, monolayer is felt implement the function that cannot realize different or door to offer, if it cannot execute different or operation, we can put a person’s mind to assume a lot of other (more interesting) the range that uses a program to will exceed a problem – solve monolayer to feel implement ability.

Fortunately, we can increase an additional node to raise nerve network to solve the ability of the problem greatly through simple ground. This feels monolayer implement became multilayer perception implement (MLP) . Mentioning like an article place, this one is called ” conceal ” , because it and outside do not have direct interface,be. I think you can regard MLP as well-known ” black box ” , it accepts the mathematical operation of data-in, executive mystery to generate output data. Conceal a layer to be inside that black box. You cannot see it, but it is over there.

Avoid duty statement: The article is reprint an article, reprint this article purpose to depend on passing more information, the person that copyright puts in original work ‘s charge is all. If involve work copyright issue,article place uses video, picture, written language, contact please small make up undertake handling.