CN108073985A

CN108073985A - A kind of importing ultra-deep study method for voice recognition of artificial intelligence

Info

Publication number: CN108073985A
Application number: CN201611034336.5A
Authority: CN
Inventors: 张素菁
Original assignee: Individual
Current assignee: Individual
Priority date: 2016-11-14
Filing date: 2016-11-14
Publication date: 2018-05-25

Abstract

The present invention relates to the ultra-deep study method for voice recognition of a kind of importing artificial intelligence in field of information processing, feature is as follows：The process micro-machine study of voice signal generates characteristic value characteristic information or voice status information is input to the input layer of ultra-deep learning neural network；Input layer is input to nervous layer by micro-machine study；Nervous layer is generated nerve signal on the basis of threshold values and is input to brains layer, and the judgement of result is identified in brains layer.Implementation result of the present invention is：Using probability scale as the threshold values of triggering neuron, with the mechanism of actual cerebral neuron very close to, it can be achieved that the really processing of artificial head brain neuron, can solve the problems, such as that complexity is as speech recognition, and computation complexity is O², learning process is with clearly defined objective, and treatment effeciency is high, this will have breakthrough on neural network theory.

Description

A kind of importing ultra-deep study method for voice recognition of artificial intelligence

【Technical field】

It is especially a kind of to import the ultra-deep study of artificial intelligence for speech recognition the invention belongs to field of information processing Method.

【Background technology】

Currently artificial intelligence becomes much-talked-about topic in worldwide, also induces one to note with the relevant patent of artificial intelligence Mesh, in this respect Japanese famous Furukawa mechanical ＆ electrical corporation delivered " image processing method and image processing apparatus " (patent document 1) Patent application, which proposes through the processing threshold values of the algorithm picks image of the neutral net of artificial intelligence so as to high-precision The profile of image is extracted out.

Famous Toyota Company of Japan has delivered the special of " drive and be directed toward estimating device " in the application of automatic driving Sharp (patent document 2), during which proposes according to automatic driving, for the situation of burst, even if driver does not have In the case of reflection, by the machine learning algorithm of the inverse transmission network of artificial intelligence, driving condition is automatically selected, with Avoid generation of driving accident etc..

【Patent document】

【Patent document 1】(special open 2013-109762)

【Patent document 1】(special open 2008-225923)

Above-mentioned (patent document 1) and (patent document 2) all mention the neural network algorithm using artificial intelligence, still, god Through the weighted value W in network algorithm, with threshold values T during study, to obtain optimal solution, it is necessary to by all states It will be tested, the total degree to be combined is { (W × T)ⁿ} × P, n is the number of nodes of one layer of neutral net here, and P is The number of plies of neutral net, the computation complexity of such high index make calculation amount huge, and it is too slow to ultimately cause self-organizing convergence；Again plus Upper weighted value and threshold values the two parameters are mutual correlations, and the adjusting of each weighted value and threshold values is carried out not for whole target It can guarantee obtained the result is that whole optimum solution；In addition, the definition of the threshold values in the model of neutral net belongs to elementary number It learns, the mechanism with the neutral net of the brain of people is very different, and the principle of the stimulus signal of cranial nerve cannot be in traditional nerve It is fully demonstrated in network model, excitement degree caused by the nerve signal according to neuron of the brains of people is different to carry out difference The mechanism of judgement can not be embodied in the model of current neutral net, then it is often random distribution to have object function, There is no consider that current neural network model can only be academicly for processing of stochastic variable etc. for neural network model , a kind of theory of directionality is represented, it is very big with the degree gap for reaching practical application.Nowadays the stage of deep learning is entered, The quantity of hidden layer is merely added compared with traditional neutral net, this more increases the complexity of calculating, although learning In imported some optimization algorithms, but and without departing from the basis of original neutral net, the fatal problem of traditional neural network It cannot solve, widely applied prospect is difficult to expect.

Further more, combinatorial theory solves the problems, such as that optimal combination is taught by Fla. university Liu by graph theory Award invention, the beginning of the eighties China visiting scholar Wang professor propose utilization " moisture in the soil " optimal combination it is theoretical, the theory due to from It can theoretically prove that optimal combined result can be obtained, therefore cause the great attention of educational circles of the world.However, utilize " moisture in the soil " Optimal combination it is theoretical the problem of be also that computation complexity is big, convergence causes slowly using being limited to.

The definition of artificial intelligence

What is artificial intelligenceIt is exactly briefly the brains function that people is realized with computer, i.e., people is realized by computer Brains thinking caused by effect, intelligent algorithm problem to be dealt with and treated the result is that unpredictable 's.

At present why socially common pattern-recognition, robot technology is confused in artificial intelligence, basic former Because being exactly unclear to the concept of artificial intelligence, therefore all advanced technologies are completely belonged to artificial intelligence, this instead can Influence the development of artificial intelligence.

People get used to the system of importing computer disposal to be referred to as intelligence system for a long time, so seeing artificial intelligence Intelligence system is just associated at once during the vocabulary of energy, this is two entirely different concepts in fact, and intelligence system is according to definite The system that the algorithm of property is realized, is the processing that certain object function is realized according to a kind of algorithm, being to determine property of handling result System.Such as automatic control system, it is adjusted by the PID of closed loop, mechanical location is enable to reach the position required in advance as early as possible It puts, temperature is made to reach prior requirement index etc. as early as possible, this algorithm is often classical theory, also in pattern-recognition There is the algorithm of the classification of many classics in intelligence system, such as using Euclidean distance, a feature vector can be calculated Data are closest with that vector data in several vector datas, these are all the rudimentary algorithms of pattern-recognition, import this The pattern recognition system of a little algorithms is exactly the system of an intelligence.

Have again in robot system, the walking of robot and the action of arm need artificially prior defeated by program Enter into robot system, the machine talent can be walked according to the program artificially inputted and various arm actions, not in the know The action that people looks at robot is also thought as that same people arbitrary can equally make various actions, and actually this is not so, in machine If a unpredictable barrier occurs in road in people's walking process, at this moment robot is certain to be tripped, however if The algorithm of artificial intelligence is equipped in robot system, it is possible to by the judgement of robot oneself, autonomous cut-through object. So the difference of common intelligence system and artificial intelligence is summed up and is exactly：Common intelligence system is classical algorithm, is only It is to solve the problems, such as the algorithm the result is that predictability to meet the algorithm of object function, artificial intelligence is imitated at brain The method of reason problem can objectively realize processing achieved by human brain, the result of problem to be solved and processing It is often probabilistic or perhaps unpredictable in advance.

【The content of the invention】

The present invention first purpose be：It is proposed the side of the speech recognition of the ultra-deep study of an importing artificial intelligence Method, this method meet the definition method of the neutral net threshold values of the nerve signal of nerve triggering caused by the cerebral nerve of people, Can adapt to neural network algorithm is application of the problem as object using the complexity of probability distribution as speech recognition.

Second object of the present invention is：It is proposed that one adapts to speech recognition practical application needs, has and calculates complexity Spend low, efficient, the algorithm of the speech recognition suitable for importing ultra-deep study.

To solve the problems of above-mentioned traditional neural network algorithm, a kind of ultra-deep study mould of artificial intelligence is proposed The audio recognition method of type, the following technical solutions are proposed by the present invention：

A kind of importing ultra-deep study method for voice recognition of artificial intelligence, feature are as follows：

Characteristic information or voice status information of the voice signal by micro-machine study generation characteristic value are input to ultra-deep The input layer of learning neural network；Input layer is input to hidden layer i.e. nervous layer by micro-machine study；Nervous layer using threshold values as Benchmark generates nerve signal and is input to output layer i.e. brains layer, and the judgement of result is identified in brains layer.

Moreover, on the basis of the study of above-mentioned micro-machine refers to probability scale or estimates scale, it is continuous by way of iteration The self-organized algorithm for generating more accurate new probability scale or the central value for estimating scale and data.

Above-mentioned probability scale refers to that a scale can be found in the data with probability distribution, can be in nominal data It is maximally distributed the scope of probability.

It is above-mentioned to estimate scale and refer to that a scale is found in the data containing probability and fuzzy message, it can demarcate Probability and the scope of most close fuzzy relation are maximally distributed in data.

Above-mentioned probability scale estimates the valve that scale can be triggered as the composition brains nerve in ultra-deep learning model Value.

The present invention also proposes that a kind of ultra-deep study of importing artificial intelligence knows method for distinguishing for image, and feature is as follows：

The process micro-machine study of picture signal generates characteristic value characteristic information, is input to ultra-deep learning neural network Input layer；Input layer is input to hidden layer i.e. nervous layer by micro-machine study；Nervous layer generates nerve letter on the basis of threshold values Number output layer i.e. brains floor is input to, the judgement of result is identified in brains layer.

The present invention also proposes a kind of constructive method of the ultra-deep learning model of artificial intelligence, and feature is as follows：

The information of several inputs of the object function corresponding to each node of the input layer of neutral net is to pass through Each node of input layer is connected to after corresponding micro-machine study processing；Between input layer and each node of hidden layer And be connected with each other by micro-machine study；The processing of handling result and self study is judged by output layer.

The present invention proposes a kind of method for the speech recognition for importing ultra-deep study, it can be achieved that really emulation brains is refreshing Processing method through member is realized and its algorithm of machine learning.The complicated system that speech recognition, which can be handled, has probability distribution is asked Topic.

The computation complexity of brand-new " ultra-deep study " is O², common intelligent terminal can be carried out applying, learn Journey is with clearly defined objective, and treatment effeciency is high, especially suitable for the realization of hardware circuit, may make up machine learning chip.After self-organizing Probability scale as the threshold values of triggering neuron and the mechanism of actual cerebral neuron very close to, and be adaptable to containing The application of the speech recognition object of random component, this will have breakthrough on neural network theory.

Description of the drawings

Fig. 1 is a kind of composition schematic diagram of the ultra-deep learning model of artificial intelligence

Fig. 2 is the composition schematic diagram of the ultra-deep learning model of actual artificial intelligence

Fig. 3 is the schematic diagram that a kind of model of ultra-deep study for speech recognition is formed

Fig. 4 is the schematic diagram that the model of the actual ultra-deep study for speech recognition is formed

Fig. 5 is the self-organized algorithm flow chart of probability scale

Fig. 6 is the schematic diagram of the building method for the Prediction of Stock Index platform for importing the ultra-deep theories of learning of artificial intelligence

In Fig. 1：

F^z(z=1,2 ..., w) is the z times by study image

F^sIt is identified image S

ML^z _1h(z=1,2 ..., w, h=1,2 ..., k) is that the information for the h image-regions for being directed to the z times image is carried out Micro-machine study

ML_2h(h=1,2 ..., k) is that input layer is carried out between hiding node layer for h input nodes information Micro-machine study

ML^s _1h(h=1,2 ..., k) is the micro-machine that the information of identified object S is carried out to information between input layer It practises

L^z _1h(h=1,2 ..., k, z=1,2 ..., w) is micro-machine study ML^z _1hIt is input to the study numerical value of input layer

L_2h(h=1,2 ..., k) is that w study of input layer micro-machine on node h learns ML_2hThe learning value of generation

T_2h(h=1,2 ..., k) is the threshold values for triggering h nervous layers

P^z(z=1,2 ..., w) is the input layer of the z times study

In Fig. 2：

F is the image or identified image learnt

ML_1h(z=1,2 ..., w, h=1,2 ..., k) be for image F h image-regions information carried out it is micro- Machine learning

L_1h(h=1,2 ..., k, z=1,2 ..., w) is micro-machine study ML_1hIt is input to the study numerical value of input layer

L_2h(h=1,2 ..., k) is that w study of input layer micro-machine on node h learns ML_2hThe learning value of generation is The learning value being input on the h nodes of nervous layer

T_2h(h=1,2 ..., k) is the threshold values for the h nodes for triggering nervous layer

In Fig. 3：

V^zVoice signal when (z=1,2 ..., w) is the Z times study

V^sIt is identified voice signal

FFT is Fast Fourier Transform

ML^z _1h(z=1,2 ..., w, h=1,2 ..., k) is the h frequency spectrums for the voice signal of the z times study The micro-machine study of habit

ML_2hIt is the micro-machine study learnt for w information of the h nodes of w input layer

ML^s _1h(h=1,2 ..., k) is the micro-machine study learnt for h frequency spectrum of identified voice signal

N^z _1h(h=1,2 ..., k, z=1,2 ..., w) is the node of the input layer of the z times study

N_2h(h=1,2 ..., k) is the node of hidden layer

N^s _1h(h=1,2 ..., k) is the node of identified voice messaging input layer

L_2h(h=1,2 ..., k) is to learn ML by micro-machine_2hThe learning value produced

T_2h(h=1,2 ..., k) is to learn ML by micro-machine_2hThe threshold values of the triggering nerve produced

P^z(z=1,2 ..., w) is the input layer of the z times study

In Fig. 4：

V is the voice signal of study or identified voice signal

FFT is Fast Fourier Transform

ML_1h(h=1,2 ..., k) is the micro-machine study learnt for the spectrum of voice signal

ML_2hIt is the micro-machine study learnt for the information of each node of input layer

N_1h(h=1,2 ..., k) is the node of input layer

N_2h(h=1,2 ..., k) is the node of hidden layer

L_1h(h=1,2 ..., k) is to learn ML by micro-machine_1hThe learning value produced

P₁(z=1,2 ..., w) is input layer

P₂(z=1,2 ..., w) is hidden layer i.e. nervous layer

P₃(z=1,2 ..., w) is output layer i.e. brains layer

In Fig. 6：

F¹It is sociological information

F²It is economics information

F^wIt is historical information

ML^z _1h(h=1,2 ..., k, z=1,2 ..., w) is the micro-machine study of z-th of predictive factors, h-th of information

N^z _1h(h=1,2 ..., k, z=1,2 ..., w) is the node of z-th of predictive factors, h-th of input layer

N_2h(h=1,2 ..., k) is h-th of node of hidden layer

N₃₁It is the output layer i.e. node of brains layer

Specific embodiment

The embodiment of the present invention is further described below in conjunction with attached drawing, but embodiment of the present invention is illustrative Rather than it is limited.

Fig. 1 is a kind of composition schematic diagram of the ultra-deep learning model of artificial intelligence.

Being defined as below for alphabetical expression is carried out first：

If input the quantity of information and the quantity of input layer for h (h=1,2 ..., k), then set study number as z= 1,2 ..., w, the number of plies of neuron for p (p=1,2 ..., e).The image learnt also should be F^z(z=1,2 ..., w), it is hidden Hiding the number of nodes of layer, that is, nervous layer, correspondings with the number of nodes of input layer also for h, (h=1,2 ..., k), learn for the first time It is ML to input information to the micro-machine study learnt before input layer^z _ph, input layer p=1, so ML^z _1hHave

(formula 1)

ML¹ ₁₁, ML¹ ₁₂..., ML¹ _1k,

ML² ₁₁, ML² ₁₂..., ML² _1k,

...,

ML^w ₁₁, ML^w ₁₂..., ML^w _1k,

So the computation complexity of the micro-machine study in the study stage each time is linear, and need to carry out w times The computation complexity of study should be strictly O³。

Re-define the node N of each layer^z _ph, in input layer p=1, so N^z _1hThere is input layer：

(formula 2)

N¹ ₁₁, N¹ ₁₂..., N¹ _1k

N² ₁₁, N² ₁₂..., N² _1k

...,

N^w ₁₁, N^w ₁₂..., N^w _1k

Equally, the input layer of second study arrives to hide and is also required to the micro-machine learnt between node layer and learns For ML_ph, there was only a kind of data in this layer, so there is no the data of w times, hidden layer p=2, so ML_2hHave

(formula 3)

ML₂₁, ML₂₂..., ML_2k

Equally, there was only a kind of data in the node of hidden layer, that is, nervous layer, so there is no the data of w times, hidden layer p= 2, so N_2hThere is input layer：

(formula 4)

N₂₁, N₂₂..., N_2k

Similary decision-making level, that is, strata cerebrale p=3, and a node is only existed, so N_3hThe node for having input layer is N₃₁。

Identified object S is defined again, the micro-machine study before the information to input layer of identified object S is ML^s _phInput layer p=1, so ML^s _1hHave

(formula 5)

ML^s ₁₁, ML^s ₁₂..., ML^s _1k

Equally, the node for being identified information to the input layer of object S only has a kind of data, so there is no the data of w times, Input layer p=1, so N_1hThere is input layer：

(formula 6)

N₁₁, N₁₂..., N_1k

The input layer of identified object S learns to the micro-machine between hidden layer as ML^s _phHidden layer p=2, so ML^s _2h Have：

(formula 7)

ML^s ₂₁, ML^s ₂₂..., ML^s _2k

Here the constructive method of the neural network model of the ultra-deep study of artificial intelligence is first illustrated so that image identifies as an example, such as Shown in Fig. 1：The image that F expressions are learnt, can divide the image into the region of n × m, can generate F_ij(i=1,2 ..., n, j= 1,2 ..., m) a image, n+m=k, the node of a learning layer is corresponded to by each region of learning object, then sets each area Domain have g (g=1,2 ..., t) a pixel.

By h (h=1,2 ..., k) gray value of the g all pixels in a region is input to first time micro-machine Learn ML¹ _1hIn, learning outcome value L¹ _1hIt is sent to h node N of first input layer¹ _1h, the z times image F learnt^zK Region learns ML by k × z micro-machine^z _1h(h=1,2 ..., k, z=1,2 ..., w) after, be input to first input layer K node on, for the obtained identification image F learnt under the varying environment of w times^z(z=1,2 ..., w), can produce The micro-machine study numerical value L of raw w × h input layer^z _1h(h=1,2 ..., k, z=1,2 ..., w) i.e.：

(formula 8)

L¹ ₁₁, L¹ ₁₂..., L¹ _1k

L² ₁₁, L² ₁₂..., L² _1k

L^w ₁₁, L^w ₁₂..., L^w _1k

By each node N of w input layer^z _1hOn L^z _1hSecond of micro-machine study ML is carried out again_2h, generate learning value L_2hAnd the scale that maximum probability or maximum are estimated, using this scale as triggering nerve

Threshold values T_2h, result is sent to k node N of hidden layer (p=2)_2h(h=1,2 ..., k).

When identified image S h (h=1,2 ..., k) a region, the result by machine learning generate learning value L^S _1h, it is sent to k node N_2h, work as by calculating | L^S _1h-L_2h|≤T_2h→ " 1 ", k node N of hidden layer_2hTriggering exports " 1 " Nerve signal, otherwise | L^S _1h-L_2h| ＞ T_2h→ " 0 " impassivity signal output.

Output layer, that is, strata cerebrale, the nerve signal exported by hidden layer number, determine the excitement degree of brain, according to The number of the output " 1 " of each node of the excitement degree of neuron, i.e. hidden layer determines whether identified image S is learnt The degree of image can play the effect of brains decision-making, in addition, in specific identification, it can also be according to multiple to identified The identification of image calculates the probability of the output " 1 " of each node of hidden layer, can carry out correcting identification parameter automatically, play The effect of self study.

Here for revised recognizer there are two types of mode, a kind of is the general of the output " 1 " for each node for calculating hidden layer The output valve of each node of hidden layer is multiplied by the probability value of the node by rate, that is, sets the output valve of each node of hidden layer For NV_h(h=1,2 ..., k), the probability value of the output " 1 " of each node of hidden layer is P_h(h=1,2 ..., k), then export The excited value of layer is

Another kind is that the data of identified image are re-fed into hidden layer as learning data, relearns to obtain new Learning outcome.

It is attached between above-mentioned node by micro-machine study, can also use statistical algorithm, for example calculated flat The method of average and variance or other statistical algorithms and other algorithms of machine learning etc., reaches above-mentioned definition Ultra-deep study constructive method, all belong to the scope of the present invention within.

Fig. 2 is the composition schematic diagram of the ultra-deep learning model of actual artificial intelligence.

It is above-mentioned, the Image Acquisition by w times is employed in image study, generates F¹, F²..., F^wA image, for F^z(z=1,2 ... w) a image passes through ML^z _1h(h=1,2 ..., k, z=1,2 ... h × w micro-machine study w), Learning outcome L^z _1h(h=1,2 ..., k, z=1,2 ... w input layer P w) is sent to respectively¹, P²..., P^wEach node N^z _1h(h=1,2 ..., k, z=1,2 ... w), then ML passed through by input layer_2h(h=1,2 ..., k) study of a micro-machine generate (L_2h, T_2h) (h=1,2 ..., k).

As shown in Figure 2：The above-mentioned generated F in study¹, F²..., F^wA image iterates through corresponding F images k The k micro-machine study ML of cell_1h(h=1,2 ..., k) obtain the learning outcome L that w is returned_1h(h=1 2 ..., k) is sent respectively To each node N of input layer_1h(h=1,2 ..., k), micro-machine study ML_2h(h=1,2 ..., input terminal k) input The h nodes N of layer_1hW obtained data carry out micro-machine and learn to obtain (L_2h, T_2h) (h=1,2 ..., k) k neutral net Learning value and threshold values, the basis for estimation of the output " 1 " as nervous layer, node N_2h, work as by calculating | L^S _1h-L_2h|≤T_2h→ " 1 ", the h nodes N of nervous layer_2hTriggering exports the nerve signal of " 1 ", otherwise | L^S _1h-L_2h| ＞ T_2h→ " 0 " impassivity signal is defeated Go out.

The nerve signal that output layer, that is, strata cerebrale is exported with above-mentioned identical, by hidden layer number, determine the emerging of brain It puts forth energy degree, according to the excitement degree of neuron, i.e. the number of the output " 1 " of each node of hidden layer determines that identified image S is The no effect for being the degree of learnt image, brains decision-making being played.

Fig. 3 is the schematic diagram that a kind of model of ultra-deep study for speech recognition is formed.

As shown in Figure 3：Initially in the study of speech recognition, by z (z=1,2 ..., w) return identical content voice letter Number V^zBy Fast Fourier Transform generate fn (fn=1,2 ..., y) data of a different spectral, by adjacent per x frequency spectrum Numerical value regards a kind of numerical value of frequency spectrum as, by h=y/x obtain h (h=1,2 ..., k) the information of a input layer, then will be per x The numerical value of frequency spectrum is input in a micro-machine study, ML^z _1hA micro-machine study generates L^z _1h(h=1,2 ..., k, z=1, 2 ..., w) a learning value, the N of input layer is sent to respectively^z _1h(h=1,2 ..., k, z=1,2 ..., w) a node, will input The learning value L of w first node of layer^z ₁₁(z=1,2 ..., w) be sent to hidden layer input micro-machine study ML₂₁In, second The learning value L of node^z ₁₂(z=1,2 ..., w) be sent to hidden layer input micro-machine study ML₂₂In, and so on, kth node Learning value L^z _1k(z=1,2 ..., w) be sent to hidden layer input micro-machine study ML_2kIn, it is generated by k micro-machine study Go out k learning value L_2hAnd the threshold values T of k triggering nerve_2h, i.e. (L_2h, T_2h) (h=1,2 ..., k), result is sent to hiding K node N of layer (p=2)_2h(h=1,2 ..., k), are stored in the memory of neutral net, as identify the sound Habit value and threshold values.

In cognitive phase, as identified voice V^sVoice caused by consistent with learning state pass through quick Fourier Leaf transformation generates fn, and (fn=1,2 ..., y) data of a different spectral, regard the adjacent numerical value per x frequency spectrum as a kind of frequency The numerical value of spectrum, by h=x/y wait until h (h=1,2 ..., k) a input information, then the numerical value of every x frequency spectrum is input to one In micro-machine study, ML^s _1hA micro-machine study generates L^s _1h(h=1,2 ..., k) a learning value, are sent to input layer respectively K node N^s _1h(h=1,2 ..., k) in, k node N^s _1hK learning value L^s _1h(h=1 2 ..., k) is directly sent to hide Layer is on the node corresponding to nervous layer, is worked as on nervous layer by calculating | L^S _1h-L_2h|≤T_2h→ " 1 ", h-th of hidden layer Node N_2hTriggering exports the nerve signal of " 1 ", otherwise | L^S _1h-L_2h| ＞ T_2h→ " 0 " impassivity signal output, all nerves Node is connected with output layer, that is, brains layer on layer, and the nerve signal that nervous layer exports is directly inputted to strata cerebrale.

Output layer, that is, strata cerebrale, the nerve signal exported by hidden layer number, determine the excitement degree of brain, according to The number of the output " 1 " of each node of the excitement degree of neuron, i.e. hidden layer determines identified voice V^sWhether it is to be learnt Voice degree, the effect of brains decision-making can be played,, can also be according to multiple pair in specific identification with above-mentioned The identification of identified voice calculates the probability of the output " 1 " of each node of hidden layer, can carry out correcting identification ginseng automatically Number plays the effect of self study.

Fig. 4 is the schematic diagram that the model of the actual ultra-deep study for speech recognition is formed.

The same with Fig. 3, the model of the actual ultra-deep study for speech recognition is formed as shown in Figure 4：In speech recognition The study stage, the voice signal V of w times^zBy Fast Fourier Transform generate fn (fn=1,2 ..., y) a different spectral Data regard the adjacent numerical value per x frequency spectrum as a kind of numerical value of frequency spectrum, by h=y/x obtain h (h=1,2 ..., k) a The information of input layer, then the numerical value of every x frequency spectrum is input in a micro-machine study, k micro-machine learns ML_1hBy W study generates w × h learning value L respectively_1h(h=1,2 ..., k), are separately input to input layer P₁K node N_1h(h= 1,2 ..., k), in the node N of input layer_1hBy w learning value L caused by w study_1hIt is sent to corresponding microcomputer Device learns ML_2hInput terminal, carry out micro-machine study after, generate k learning value L_2hAnd most trigger the threshold values T of nerve_2h, i.e., (L_2h, T_2h) (h=1,2 ..., k), result is sent to k node N of hidden layer (p=2)_2h(h=1,2 ..., k), are stored in In the memory of neutral net, as the learning value and threshold values for identifying the sound.

In cognitive phase, as identified voice V^sVoice caused by consistent with learning state pass through quick Fourier Leaf transformation generates fn, and (fn=1,2 ..., y) data of a different spectral, regard the adjacent numerical value per x frequency spectrum as a kind of frequency The numerical value of spectrum, by h=x/y wait until h (h=1,2 ..., k) a input information, then the numerical value of every x frequency spectrum is input to one In micro-machine study, k micro-machine learns ML_1hGenerate k learning value L_1h(h=1,2 ..., k), are sent to input layer respectively K node N^s _1h(h=1,2 ..., k) in, k node N^s _1hK learning value L_1h(h=1 2 ..., k) is directly sent to hidden Hide on the node corresponding to layer, that is, nervous layer, work as on nervous layer by calculating | L^S _1h-L_2h|≤T_2h→ " 1 ", the h of hidden layer A node N_2hTriggering exports the nerve signal of " 1 ", otherwise | L^S _1h-L_2h| ＞ T_2h→ " 0 " impassivity signal output, all god It is connected through node on layer with output layer, that is, brains layer, the nerve signal of nervous layer output is directly inputted to strata cerebrale.

Output layer, that is, strata cerebrale, the nerve signal exported by hidden layer number, determine the excitement degree of brain, according to The number of the output " 1 " of each node of the excitement degree of neuron, i.e. hidden layer determines identified voice V^sWhether it is to be learnt Voice degree, the effect of brains decision-making can be played.

When specifically carrying out speech recognition, to improve the precision of speech recognition, it is also contemplated that by voice signal according to speaking Word or voice status, first voice signal is split, each content and state of voice are distinguished respectively Spectrum analysis, and the probability value that state changes also is included in learning Content, it can be according to several states not in identification Same combination judges the content of voice and considers the probability of voice status transformation to improve the accuracy rate of speech recognition.

Above-mentioned identified image is only to carry out region segmentation, then carries out processing and the voice of ultra-deep study Spectrum analysis as a result, not exclusively this processing method, can also artificial intervention be passed through figure according to the algorithm of space reflection Picture or voice are mapped to each simple space, then are handled by the method for above-mentioned ultra-deep study.

The principle of space reflection is described below.This theory of space reflection is to belong to the theory of artificial intelligence scope.Space The principle of mapping theory is to be directed to similar recognition of face, and the pattern-recognition of complicated system as image identification or Text region is asked Topic, cannot be as traditional intelligence system, directly since the problem of complexity is is difficult often to find the algorithm that can directly solve What is connect is handled by traditional algorithm, and the space reflection theory based on fuzzy mathematics is the problem of a complexity being space The problem of being mapped to the space of several Simple Systems, although can only be solved the problems, such as in the space of each Simple System it is limited, But can but solve the problems, such as that complexity is space according to the combination in several Simple System spaces of combinatorial theory, it is most important herein It by complexity is that space problem is mapped to Simple System space to be, is the processing of the brain based on people, academicly referred to as human world intervention Method is not easy to carry out human world intervention since traditional mathematical method is very inflexible, and fuzzy mathematics, can be with to being we provided facility According to the understanding artificially to process object, fixed pattern is carried out from several angles by the Membership functions of fuzzy mathematics, from And realize be to complexity problem solution, since such algorithm is the brains processing mode according to people again by fuzzy mathematics Fixed pattern achievees the effect that solve the problems, such as that complexity is, therefore should belong to the theoretical category of artificial intelligence.

Very high-caliber application effect can have been obtained in the automatic identification of handwriting digital using such theory, it can With according to artificial understanding word as a result, such as in difference digital " 9 " and the identification of number " 4 ", when the scanner used Resolution there was only 100dpi, number " 9 " after scanning and digital " 4 " are if digital " 4 " hold very much when following stroke is longer It is easy to identify into digital " 9 ", in turn hand-written " 9 " when following stroke is shorter, may be identified as digital " 4 " as a result, It, can be the fuzzy value of digital " 9 " and digital " 4 " using the Membership function fixed patterns in space reflection this theory It can quantify, it is possible to for the identification problem of non-online handwriting word, obtain very high-precision recognition result.

The theory of above-mentioned space reflection is imported, shape of the image on different position, direction can be directed to, size is concentrated It is scattered, geometric feature and center of gravity including being uniformly distributed etc., energy, heat, frequency spectrum, density, derivative, texture, gray scale The characteristic of the physics such as the ratio of value adds in artificial subjective analysis, according to artificial understanding and artificial discriminating conduct structure Membership functions are built, form the model of each space reflection to being entirely identified image, formation can express characteristics of image Each input data.

The probability scale self-organization theory that proposition is mutually competed with traditional nerual network technique is to belong to artificial intelligence theory. This theoretical starting point is in order at artificial thought, can pass through self-organizing if the scale of a most probable value can be found Method obtain a beyond tradition probability value maximum solution, thereby produce the algorithm of probability scale self-organizing.

Beginning of the program making person of traditional algorithm to program, processing procedure that is intermediate or even terminating all is that programmer is prior It designs, there is foreseeability, and a distinguishing feature of probability scale self-organized algorithm is exactly programmer to processing Process and result all with unpredictability.

Before this theoretical appearance, the handling result of all algorithms related with statistics all rests on this Before the processing of algorithm, the conversely place for the algorithm that the result after the algorithm process can allow all related with statistics It manages result and generates breakthrough, and this algorithm occurs being considered as impassable statistical various constants by people at present afterwards Through not being optimal constant.

Compared with the algorithm of deep learning, the target of the self-organizing of probability scale self-organized algorithm is clear, ultrahigh in efficiency, Iteration must have effect each time, and computation complexity is linear, and regular handset APP is achieved that, great application prospect.

Due to the breakthrough of this algorithm in theory, computation complexity is low so that its appearance shows that its is special always Application effect, such as on Text region OCR system, on the document files that goes out in computer printout, in no datum line In the case of, when file puts inclined on the scanner using this algorithm, only by the arrangement of word can be quickly calculate File puts inclined angle.

In the application of recognition of face, such as the position of face is found, traditional method is first to provide the number of colours of face According to the method for according to program touch melon along Teng finds and belongs to all pixels of face color, and problem is a similary people not With light under the color distinction of image shot it is very big, then have in the world and have the different colours of skin, the tool that a kind of colour of skin is included Body color is also the need that a kind of multifarious, traditional algorithm of color into line search of definition cannot meet practical application certainly Will, import probability scale self-organized algorithm directly only can accurately can find face by self-organizing several times Position, because regardless of the colour of skin, no matter due to the cross-color of the image captured by the difference of shooting light, in whole image In face position color distributed density values it is maximum, that is to say, that the probability value of the colour of skin at face position is maximum, by general The algorithm of rate scale self-organizing can simply solve this problem very much certainly, and without the method for touching melon along Teng, first Can face position automatically be shifted to during probability scale self-organizing by any position of image during the beginning, and it is final The profile at entire face position is provided, this is not imaginable recognition effect in the algorithm of traditional pattern-recognition.So one A algorithm for surpassing deep learning can instantaneously realize that probability scale self-organization theory should belong to only by mobile phone terminal Machine Learning Theory.

The micro-machine study used in a kind of ultra-deep learning model of artificial intelligence of the present invention is introduced in detail below, it is actual Upper is exactly the algorithm of above-mentioned probability scale self-organizing, and the algorithm learnt between the node of neutral net is knitted using micro-machine is carried out Connection can make neural network theory generate breakthrough development, specifically introduce the micro-machine for importing probability scale self-organizing below The specific algorithm of study.

If the given one ordered series of numbers g with probability distribution₁, g₂... g_ζCollection be combined into G ∈ g_f(f=1,2 ..., ζ), should The central value of set is A (G), and the probability scale that central value is A (G) is M [G, A (G)], and is calculated by self-organization iteration With the central value A (G of (n-1)th time^(n-1)), and the radius M [G on the basis of the central value^(n-1), A (G^(n-1))] in there is The ordered series of numbers g of k probability distribution₁, g₂... g_kCollection a be combined into G⁽ⁿ⁾∈g_f(f=1,2 ..., k), then

(formula 9)

A⁽ⁿ⁾=A (G⁽ⁿ⁾)

M⁽ⁿ⁾=M [G⁽ⁿ⁾, A (G⁽ⁿ⁾)]

G⁽ⁿ⁾=G { [A (G^(n-1)), M [G^(n-1), A (G^(n-1))]]

It is the ordered series of numbers g for probability distribution to pass through the central value that iteration is calculated several times by above-mentioned iterative formula 1₁, g₂... g_ζThe estimate of obtained closest parent, and final horizon radius value is a probability scale, in final On the basis of center value, the ordered series of numbers g of all probability distribution in the range of probability scale₁', g₂' ... g_k' it can each belong to probability point Cloth ordered series of numbers g₁, g₂... g_ζTrue value.

Fig. 5 is the self-organizing process flow of probability scale distance.

As shown in Figure 5：If the given one ordered series of numbers g with probability distribution₁, g₂... g₁Collection be combined into G ∈ g_f(f=1, 2 ..., 1), then it is made of based on the self-organized algorithm of probability scale following 4 steps.

STEP1：Pre-treatment step：M⁽⁰⁾As initialization probability yardstick, A⁽⁰⁾As the initial centered value of self-organizing, V makees For the convergency value of self-organizing, MN is as self-organizing maximum tissue time numerical value, current numbers of the initial n=0 as self-organizing.

On M⁽⁰⁾As initialization probability yardstick and A⁽⁰⁾The determining method of initial centered value as self-organizing, without into The strict setting of row.By manual prediction, for final scope, at least part of numerical value is included in initialization probability ruler Spend M⁽⁰⁾In the range of.Initialization probability yardstick M⁽⁰⁾Bigger, the time of calculating is longer, otherwise too small, it is possible to cannot Correct result.

On setting methods of the V as convergency value, convergency value V is bigger, it is possible to cannot get correct result.Convergency value Smaller, the time for calculating cost is longer.Correct setting method is 10% or so of the probability scale of final self-organizing.

On the setting method of the maximum number of self-organizing MN, it is sufficient to be usually 5-10 times.

STEP2：Self-organizing step：N times self-organizing processing is carried out, A⁽ⁿ⁾As self-organizing central value, probability scale M⁽ⁿ⁾ As radius, with central value A⁽ⁿ⁾On the basis of, calculate all numerical value g within radius_fThe average value V of (f=1,2 ..., ζ)⁽ⁿ⁺¹⁾ With dispersion value S⁽ⁿ⁺¹⁾, V⁽ⁿ⁺¹⁾=A⁽ⁿ⁺¹⁾, S⁽ⁿ⁺¹⁾=M⁽ⁿ⁺¹⁾, n=n+1.

(formula 10)

(formula 11)

STEP3：Self-organizing discriminating step.Self-organizing processing reaches maximum times (N >=MN) or self-organizing processing convergence (M⁽ⁿ⁾-M⁽ⁿ⁺¹⁾≤ V), it is such as YES, the self-organizing for no longer carrying out next time is handled, self-organizing terminates to jump to STEP4.If NO just jumps to STEP2 and continues self-organizing processing.

STEP4：Self-organizing processing terminates.

Probability scale M⁽ⁿ⁾It is the parameter of a probability statistics with multiple attributes.For example normal distribution, index point Cloth, Erlangian distribution, Weibull distribution, angular distribution, beta distribution etc..Such as probability scale M⁽ⁿ⁾It can serve as normal state point The dispersion value of cloth.

The algorithm of probability scale self-organizing makes traditional statistical basic constant generate subversive breakthrough, and basic The closely related correlation analysis of constant, regression analysis etc. will all generate breakthrough, and new statistical algorithm will generate.It will with this The relevant theory of new statistics being formed under artificial intelligence theory.

Scale problem in machine learning, the algorithm of probability scale self-organizing so as to generating breakthrough result and being exactly Because probability scale and the computational methods of self-organizing are introduced.That summarizes current all traditional related scales defines picture More not significantly more efficient than probability scale, still, whether the algorithm of probability scale self-organizing can not develop with regard to thisIt answers It is negative, the fuzzy item probability of above-mentioned solution optimization combinatorial theory is actually a to be estimated, and the scale estimated is believed can To generate new breakthrough, because estimating scale not only letter of guarantee probabilistic information, while letter of guarantee fuzzy message is gone back, and estimate and will make With unconspicuous small probabilistic information and small fuzzy message integrated to obtain it is more stable estimate, with survey Spend as self-organizing scale this by be current highest level machine learning algorithm, this is also the artificial intelligence that people predict Basic theory be combinatorial theory since estimating the optimization problem that can be solved in combinatorial theory as described above, just can completely Proof estimates the ideal algorithm that scale self-organized algorithm is machine learning.This is also probability theory, fuzzy theory with from The theory of organization has gone to the ultimate stage in informatics, it means that will generate important breakthrough, people in message area Work intelligence also rests on initial stage always by lacking the support of basic theory only with rule, moves towards with basic theory For the advanced stage of support.

It is theoretical that the optimal combination based on fuzzy item probability is set forth below, combinatorial theory is the basic theory of artificial intelligence, Therefore the breakthrough of artificial intelligence theory is necessarily dependent on the breakthrough of combinatorial theory.

Combinatorial theory problem to be solved is how efficient to realize optimal combined result and in face of extensive Integrated circuit needs to realize that area is minimum, and most short or even be also contemplated that the combination of more purposes such as electrical characteristic with line length, this is to pass The insurmountable problem of combinatorial theory institute of system, proposes that the optimal combination of a fuzzy item probability is theoretical, this is theoretical here Connection relation between the modules of complicated integrated circuit is estimated carry out quantification by fuzzy item probability, is passed through Consider the fuzzy relation that the connection relation between unit is more close more to be arranged together as far as possible, while be also contemplated that pin To the probabilistic relation that a unit is possible to arrange near this unit with unit, unconspicuous small probability is believed Breath and unconspicuous small fuzzy message integration get up can be obtained by stable, it will be apparent that and valuable information, Here it is the break-through point of fuzzy item probability theory, therefore can be efficient, the optimization for the integrated circuit of more purposes needs It asks, directly calculates optimized combined result.This theoretical basis be in order at artificial subjectivity to the company between unit The definition of the fuzzy value of relation is connect, therefore is also the scope for belonging to artificial intelligence theory.

Belong to the integration factor of 40 or more as combinatorial theory and belong to the unsolvable np problem of Turing machine, for playing chess Problem is to belong to np problem, and still, if adding in artificial experience, np problem still can solve, if the experience played chess is done It is suffered to program, therefore can realize the effect for defeating chess player.

Here to the definition of fuzzy phenomenon probability measure.Probability measure (Probability Measure) is defined first：Letter Probability measures of the number μ on probability space must is fulfilled for empty set and is combined into " 0 ", and complete or collected works are combined into " 1 ", if set A is set B A part, then the probability value of set A is less than the probability value (monotonicity) of set B and meets computability.

Re-define fuzzy mearue (Fuzzy Measure)：Fuzzy mearue of the function mu on fuzzy space must is fulfilled for empty set It is combined into " 0 ", complete or collected works are combined into " 1 ", if set A is a part of set B, then the fuzzy value of set A is fuzzy less than set B It is worth (monotonicity), still, fuzzy mearue is unsatisfactory for computability.

The definition of fuzzy phenomenon probability measure (Probability Measures of Fuzzy Events)：It assumes first that Function f (x) is independent in the set of fuzzy space, therefore with computability, f (x) and p (x) be divided into do not meet it is above-mentioned The condition of probability measure and fuzzy mearue, then obscuring phenomenon probability measure is：(formula 12)

The physical significance of formula 12 is：The problem of being for a complexity, can be by by numerous small fuzzy letters Breath utilizes their measure properties with small probabilistic information, can obtain a stable value of information by the method for integration, The unexpected effect for solving the problems, such as complexity and being can be obtained using this fuzzy probability measurement as the scale of self-organizing.

The present invention is described below can carry out following concrete application using above-mentioned theory：

First target identification is moved in 3 D stereo

Intelligent algorithm has been imported in mobile object identification technology, can be directed to since rugged environment causes image In the case of severe jamming, mobile target can be accurately hit very much.

Currently by mapping of the UAV to topography and geomorphology, wrecked personnel etc. are searched for automatically and are required for having manually The identification product of the three-dimensional mobile object of intelligence, such high-end technology industry have unlimited commercial value.

Automatic stock exchange fund is described below to liquidate and the application process of financial prediction.

Socially most valuable technology is Predicting Technique, because correctly prediction stock market can obtain huge wealth Richness, the needs of where you are and where it's at you see the correct algorithm for estimating stock market's result can not but meet people, therefore, in Forecasting Methodology Even there is some technological progresses that will all be very important, some of the U.S. are known as the prediction algorithm of military secrecy, after disclosure Even if we can feel these algorithms, at that time, there is no that high-caliber technological progresses among imagining.

But entering the epoch of artificial intelligence, Predicting Technique will show prominent subversive effect, first optimal pre- Conceptive display its progressive surveyed, previous people thirst for obtaining an optimal predicted value, however, according to mathematically most The theory of good neutralizing, optimal neutralizing are centainly built upon the optimization on a certain boundary condition provided.Artificial intelligence it is optimal Change predicted value and be just built upon understanding of the dopester to social factors, to predicting the awareness of target and the intelligence of individual Obtained optimization value under the influence of power situation etc. factors and all conditions, this will push prediction theory to most high-order Section.

First artificial intelligence thus can optimize prediction on have it is breakthrough, one is used probability self-organizing manage By having overturned traditional statistical prediction.The second is used the space reflection of fuzzy mathematics theoretical, it can be dopester couple Social factors and the understanding of relation of prediction object are built into social model by Membership function fixed patterns, can be having The sociology to tell on to prediction is closed, philosophy, history even Eight Diagrams, that is, middle extracted experience easy to learn is in artificial intelligence Optimization forecasting system in can fixed pattern, can act as to optimize prediction effect.The third is in mathematical model The Linear correlative analysis of time series can be imported, function approaches, regression analysis, least square method, the increment of biology with it is dead It dies and various statistical models and constant, all one of element as prediction.The fourth is it can establish social Expert system establishes large-scale social accomplished expert storehouse for prediction object.Finally, the ultra-deep study of artificial intelligence is passed through Model foundation optimize predicting platform, by all algorithms related with prediction, knowledge and information are all of getting up.

Artificial intelligence, which optimizes Prediction of Stock Index platform, to be calculated these algorithms are separated, but is built into one Optimize prediction policy of Central Government platform, the result of calculation of a variety of algorithms is merged, is mutually authenticated, information that This is shared, and passes through machine learning algorithm and finally carry out self-organizing operation, eliminates the false and retains the true to obtain and surmounts statistical calculating As a result, draw the predicted value of maximum probability.System is also equipped with being drawn with each algorithm of the system with the data having occurred and that Result carry out the function of automatic Evaluation, the various parameters of automatic update the system balance the actual effect of the influence of various factors Fruit, realize automatically the renewal of knowledge and knowledge accumulation, these be all automatically on the basis of moment realize, in this platform Upper prediction result includes stock exchange, and it is all automatically to carry out that fund, which liquidates,.From another point of view, as system, it is still necessary to artificially handle Function, during operation can constantly according to raising of the operator to the understanding of predictive factors, artificial amendment various factors Numerical value increases information, increases component of forecast or again or prediction level is continuously improved in frame etc. of adjustment prediction strategy. Such system can the correctly predicted brains for being derived from people wisdom, but be mankind's prestige in high speed processing crisp decision making Dirt not and, such system be bound to liquidate in automatic stock exchange fund and financial prediction on play immeasurable work With.

Fig. 6 is the schematic diagram of the building method for the Prediction of Stock Index platform for importing the ultra-deep theories of learning of artificial intelligence.

As shown in Figure 6：It will be with predicting relevant sociological information F¹In possessed fn (fn=1,2 ..., y) a factor Classification, it is assumed that have x factor per one kind, artificial intervention is carried out by the Membership functions of fuzzy mathematics, constructs h=y/ X, that is, h (h=1,2 ..., k) the information of a input layer, then the numerical value of every x factor is input to a micro-machine study M¹ _1h(h =1,2 ..., k) in, in the same way by information F in the economics related with prediction²And it related with prediction goes through Information F in history^wEach factor learnt by numerical value caused by artificial intervention also by micro-machine, such w × h microcomputer Device learns ML^z _1hGenerate w × h learning value L^z _1h(h=1,2 ..., k, z=1,2 ..., w), are sent to input layer P respectively_z(z= 1,2 ..., N w)^z _1h(h=1,2 ..., k, z=1,2 ..., w) a node, for the learning value of each input layer L¹ _1h, L² _1h..., L^w _1h(h=1,2 ..., k) again with micro-machine learn ML₂₁, ML₂₂..., ML_2kStudy obtains L₂₁, L₂₂..., L_2kA learning value and T₂₁, T₂₂..., T_2kA threshold values, each learning value is reflected is produced in this info class Raw predicted value, predicted value and threshold values are sent to brains layer by each node of nervous layer and carry out decision-making, the handle in brains layer Estimation range is calculated in predicted value caused by all factors, and hypothesis testing is carried out in this scope, it is assumed that inspection Mode is such, and brains layer provides a test value TV_i(i=1,2 ..., r), r is the number examined, by TV_iIt is dealt into nerve On each node of layer, carried out on each node of nervous layer | TV_i-L_2z|≤T_2z→ 1, | TV_i-L_2z| ＞ T_2z→ 0, (z= 1,2 ..., w, i=1,2 ..., r), with above-mentioned identical, brains layer finds out maximum excitement according to the stirring conditions of nervous layer Predicted value, using this value as final predicted value.If prediction of failure, brains layer can also go to evaluate and test with the value actually generated Whether the fuzzy parameter set by each factor is correct, adjustment fuzzy parameter that can be automatic or artificial, reaches full accuracy Prediction effect.

In conclusion a kind of Prediction of Stock Index method of the ultra-deep study of artificial intelligence, it is characterized in that will be related with prediction The input layer of ultra-deep study is sent in the processing that all information or the result of calculation of mathematical forecasting model are learnt by micro-machine Each node；Each nodal information of each input layer learns to generate predicted value by micro-machine and threshold values is sent to nervous layer； Brains layer draws estimation range according to each predicted value, and proposes that test value arrives prediction of each nervous layer by the nervous layer again Value or threshold values carry out excited inspection, finally draw optimal predicted value.

On the basis of the study of above-mentioned micro-machine refers to probability scale or estimates scale, constantly produced by way of iteration The self-organized algorithm of raw more accurate new probability scale or the central value for estimating scale and data.Above-mentioned probability scale Refer to that a scale can be found in the data with probability distribution, it can be with the scope of maximum distribution probability in nominal data.

The above-mentioned all information related with prediction include sociological information, economics information, philosophy information, law science letter Breath, finance information, political science information, historical information, stock information, fund liquidate information, predict relevant historical data letter At least one of breath information.

The result of calculation of above-mentioned mathematical forecasting model is included through the obtained predicted value of correlation analysis computational methods, curve The obtained predicted value of approximation computation method, the predicted value that statistical models are obtained, the predicted value that least square method is obtained, What at least one mathematical model including the predicted value that regression analysis is obtained was obtained obtains result of calculation.

The application process in terms of automatic driving is introduced again.

The automatic vehicle control system for importing artificial intelligence theory is the application problem that current industry circle is paid close attention to the most.At this One is importing the pattern recognition system of the machine Learning Theory of artificial intelligence in a application field, road conditions can be believed online Breath automatically identifies, for foundation of the automated driving system as automobilism, the second is the automobile for importing artificial intelligence is automatic Why runtime, automatic driving need artificial intelligence, and exemplified by the control that can brake, automobile can not possibly be with one first Speed is run, and has a lot of situations when needing and stop on certain position, experienced driver does not touch on the brake directly be parked in sometimes The position needed, lightly steps on brake sometimes, it is also possible to and brake etc. is stepped on strength, has many situations, Such control problem is that traditional Theories of Automatic Control all at present can not solve, and imports the fuzzy deduction skill of artificial intelligence Art can the experience of experienced driver by Membership function fixed patterns, according still further to fuzzy deduction algorithm realize it is same The close automatic Pilot control of experienced driver.

Here what is enumerated is only brake control, and the automatic Pilot on real road also has more complicated control problem, Therefore it is imperative to import intelligent algorithm.

Related ITC images transform code application process.

With the evolution of code technique, had developed to now without designing code sign in advance, form code pattern with It asks to obtain stable recognition result.Under the algorithm of artificial intelligence, according to natural paper line, vocal print, natural image is even raw Body information can be directly converted to code.

In recent years the upper popular AR technologies of society can pass through some printing image of mobile phone photograph, it is possible to online connection A certain website.Since this technology can download program of increasing income from network, so rapid proliferation.But AR technologies are to pass through figure As the algorithm of identification, recognition result is the file of more than ten000000 memories of occupancy, is unfavorable for network operation and substantial amounts of image Application.

From another angle, Google glass, image retrieval be required for directly surfing the Internet by shooting an image or into Row network retrieval.The implementation method of the ITC (Image To Code) of the direct transform code of image is proposed therefore, is reflected with space The algorithm penetrated can be some latent structures of image into the feature vector of image, then the algorithm groups by probability scale self-organizing It is made into one 10³⁶Code.As soon as a code can be become by any one image by the shooting of mobile terminal by realizing, It is to say any image can be used directly as Quick Response Code, this achievement can allow any commodity sign without any processing In the case of become a Quick Response Code, all products in the world can be made, may be connected among a night and go on the net, no Destroy the beauty of commodity sign.It can realize that Google glass sees that any image can connect the imagination of network, can realize The development of online retailing can be promoted directly in the internet retrieval commodity by any commodity image of mobile phone photograph, then have for The development of current VR products will play an important role etc..There is code capacity compared with traditional AR in ten a ten thousandths, It is identified convenient for mobile phone terminal, occupies the characteristics of server capacity is small, and retrieval rate is fast, suitable for international a wide range of, large capacity Application.

Claims

1. a kind of importing ultra-deep study method for voice recognition of artificial intelligence, feature are as follows：

Voice signal generates the characteristic information of characteristic value by micro-machine study or voice status information is input to ultra-deep study The input layer of neutral net；Input layer is input to hidden layer i.e. nervous layer by micro-machine study；Nervous layer is on the basis of threshold values It generates nerve signal and is input to output layer i.e. brains layer, the judgement of result is identified in brains layer.

2. a kind of importing ultra-deep study method for voice recognition of artificial intelligence according to claim 1, feature It is：On the basis of the study of above-mentioned micro-machine refers to probability scale or estimates scale, constantly generated more by way of iteration Add the self-organized algorithm of accurate new probability scale or the central value for estimating scale and data.

3. the ultra-deep study method for voice recognition of a kind of importing artificial intelligence according to claim 1 or 2, It is characterized in that：Above-mentioned probability scale refers to that a scale can be found in the data with probability distribution, can be with nominal data The scope of middle maximum distribution probability.

4. the ultra-deep study method for voice recognition of a kind of importing artificial intelligence according to claim 1 or 2, It is characterized in that：It is above-mentioned to estimate scale and refer to that a scale, Ke Yibiao are found in the data containing probability and fuzzy message Fixed number is maximally distributed probability and the scope of most close fuzzy relation in.

5. according to claim 1, the ultra-deep study of a kind of importing artificial intelligence described in 2,3 or even 4 is used for the side of speech recognition Method, it is characterised in that：Above-mentioned probability scale or estimate scale can as in ultra-deep learning model composition brains nerve triggering Threshold values.

6. a kind of ultra-deep study of importing artificial intelligence knows method for distinguishing for image, feature is as follows：

Picture signal generates the characteristic information of characteristic value by micro-machine study, is input to the input of ultra-deep learning neural network Layer；Input layer is input to hidden layer i.e. nervous layer by micro-machine study；It is defeated that nervous layer generates nerve signal on the basis of threshold values Enter to output layer i.e. brains layer, the judgement of result is identified in brains layer.

7. a kind of constructive method of the ultra-deep learning model of artificial intelligence, feature are as follows：

Information of several inputs of object function corresponding to each node of the input layer of neutral net be by pair Each node of input layer is connected to after the micro-machine study processing answered；It is also between input layer and each node of hidden layer It is connected with each other by micro-machine study；The processing of handling result and self study is judged by output layer.