| The artificial nerve network is owing to the initial understanding of person the brain organization structure, activity mechanism to put forward of a kind of new information processing system.Pass a mimicry brain organization structure and some activity mechanism of the nervous system, the artificial nerve network can present many characteristics of a person's brain, and some basic functions of someone's brain.The structure, function and ability of the nerve network is 3 basic characteristicses.1. Structure characteristics Information processing of proceed parallelity , information to store of distributing , information processing unit of with each other allied , structure of plasticity.The artificial nerve network is in great quantities and in brief in order to handling a component to link the non- line form system that the component height proceed together mutually and have large-scale proceed parallelity processing characteristic.Although each the function which handle unit is very simple, in great quantities and in brief handle proceed together of unit an activity to make the network present a function for enrich and have quicker speed.The proceed parallelity of structure make nerve network of the information be by all means saving to adopt a distribute type method, namely information not is saving in a certain part of network, but distribute in the network all conjunction powers.A nerve network can save various informations, among them the conjunction power of each nerve dollar in saving is various a part of the informations.When need to be acquire saving knowledge already, the nerve network is input an information to encourage under adoption the way of "association of thought" carry on recollection, as a result have association of thought memory function.Nerve network's insidely proceeding parallelity and distributing performance is all what space ascend distribute, time ascend proceed parallelity with processings at save of its information.2. Function characteristicsHeight of not the line goodly permits wrong and compute of not precision.Nerve dollar extensively and with each other unite and proceed together a work to otherwise make the whole network present a height of not line characteristics.But the distribute type save of the structure characteristics will make the network express to permit wrong goodly in 2;On the other hand, because of the distribute type of information saving, be network central part diverted through dollar damage will not bring influence to the whole function of system.This elephant person everydays in brain have a nerve cell to normally die but will not influence cerebral function similar;On the other hand, be importation misty, incomplete or the information for transform, the nerve network function passes memory of associate the instauration integrity and carry out thus to not complete importation information of exactitude identify, this characteristics elephant person can to not the handwritten word of norm carry on exactitude to identify similar.The nerve network can handle continuous emulation signal and not accurate of the misty information of incompletion therefore to of is that the time excellently approaches solution rather than precision solution.3.ability characteristicMean that a system can tell then from the adaptability the function of oneself is with the variety ability of the orientation environment and it is an important characteristic of nerve network.Include from the study from the adaptability with from organize 2 F meaning.Nerve network of mean from the study be outside environment occurrence variety, through a period of time of training or feeling know, the nerve network can pass automatically adjust network structure parameter.Make for give exportation of settle the creation hope of the importation ability, train the path that is a nerve network study, so usually study and train two phrases to mix to use.The nervous system can stimulate in the exterior under press the certain rule adjustment the outstanding conjunction of of the nerve dollar, gradual Gou piece nerve network, this sets up process to be called a network of from organization.(or weigh Gou)Nerve network of from organizational skill with from adaptability related, pass to carry out from the organization from the adaptability.The artificial nerve network is new which draws lessons from a network at the living creature nerve and develops can information processing system, up imitated to build the living creature nervous system of person's brain because of its structure, as a result the top of its function also had a certain can characteristics.The underneath carries on synopsis introduction to the basic function of nerve network.1.Associate memoryBecause the nerve network has to distribute a saving information with proceed together calculating function, so it has outward boundary incitement an information and inputs mode to carry on the ability which associates memory.This kind of ability is the collective behavior which passes structure and information processing of the be in conjunction with of of nerve dollar but realization.The nerve network is the memory which touches power value and conjunction structure to express an information through its Tu and this kind of distribute type is saving to make the nerve network be able to save the information of more and complicated mode and instauration memory.The nerve network pass in advance saving information and study the mechanism carry on training from the orientation, can the instauration in never complete information and Zao voice interference originality of complete information, this ability makes it recover in the portrait, portrait and speech handle, the mode identify, classification etc. the aspect have huge latent application value.Associate the memory has two kinds of basic forms:Associate memory from the association of thought memory and the difference(1) Remember from the association of thoughtIn advance saving(memory) and various mode informations in the network, be input a certain have already saved parts of informations of mode or take an information of have the noise interference, the network can pass dynamic state association of thought process to start to recall all informations of that mode.(2)The different association of thought rememberIn advance saved several modes in the network to, each modes all constitute to°from two parts and be to input a certain mode one part of the rightness, even input an information is incomplete or folded to add noise, to the network can also start to recall with it to in response to of another a part.2.Not the line reflects to shootIn the objective world, the of many importations of systems and exportation exists complicatedly non- line sexual behavior, for this system, usually can hardly build up its mathematics model with the traditional mathematics method.Design the reasonable nerve network passes to input to output sample to approach arbitrarily and complicatedly and not line to reflect to shoot towards carrying on an automatic study, can with arbitrarily accuracy to the system.This good function of nerve network makes it can conduct and actions many Weis not the line mathematics model of the pass of function.The expression of that model is a cent to analyze of and the reflect of of the importation exportation data's shooting rule is been study a stage by the nerve network to automatically sample and the distribute type be saving in all conjunctions of network.Haven't line to reflect a nerve network of shoot the function is applied very vast, almost involve all realms.3.Classification with identifyNerve network outward boundary importation the sample have to very strongly identify with categorize ability.Actually is the partition district which finds out to meet a classification request in the sample space to the classification of importation sample, each sample in the district belongs to 1 type.Tradition classification method in keeping with resolve the same kind gather together, the different separate of identify with classification problem.But objective very many things(for example different portrait, voice, writing etc.) in the worlds' partitioning curved face in the district of sample space be very complicated, the close by sample may belong to a different type, but the sample of principle may belong to the same 1 type.The nerve network can nicely solve a rightness not the line approach of curved face, so have a better classification and identify ability than the traditional classification tool.4.Excellent turn a calculationExcellent turn a calculation to mean under the stipulation condition that have already know, look for a set of parameter to combine, make from should combine, make the target function which is make sure by that combination attain a minimum value.The nerve network of some types can need the appearance which requests to solve the variable parameter design of problem as network, design the target function as the energy function of network, the nerve network through dynamic state turns into process to attain stability appearance to in response to of the energy function be minimum and thus its stability appearance be the superior solution of problem.This kind of is excellent to turn a calculation not to need to beg to lead to the target function and it is a network auto as a result to.5.The knowledge handleKnowledge is people the summary induces from a great deal of information of the objective world and time of the oneself of experience, rule with judge according to.When the knowledge can carry on a description with concept and model of explicit definition, the calculator has very quick processing speed with very high operation accuracy.And under the a lot of condition, the knowledge usually can't express with explicit concept and model, perhaps the definition of concept be very misty, even the problem-solving information be not complete, not overall, for this knowledge processing problem, the nerve network acquires knowledge of path is likeness to mankind and also sample regulation but acquire knowledge concerning object from the importation of the object the exportation the information, and distribute knowledge in the conjunction of network to give saving.The knowledge of the nerve network samples ability to make it can under the sistuation that didn't check knowledge first the auto withdraws a characteristic and discovers regulation from the importation data, and pass to set up a network from the organization process, make it, suitable for expression the regulation discover.On the other hand, the person checks knowledge first and can consumedly improve the knowledge processing of nerve network ability, their combine together will make nerve network the intelligence get to promote further.All increased a momentum item in the BP calculate way currently, with as for have the BP calculate way of momentum item to become a kind of new standard calculate way.Study rate also be called to tread long, settle for the constant in the standard BP calculate way, however can hardly make sure 1 in physically appliedly from start to the eventually all suitable best study rate.Can see from the error margin curved face, too small will make the training number of times increase in the flat district, as a result hope to enlarge a value;But change violent district in the error margin, too conference because the adjustment once measures big but step over narrow"the pit be cave" place, make the training appear a vibration, on the contrary make Die generation the number of times increase.Refrain from rash action process for the sake of the acceleration, a better way of thinking is from orientation change study rate, make it should big hour enlarge, that let up.Exist flat district on the error margin curved face.The power value adjustment gets into the reason of flat area is nerve dollar the exportation got into transformation function of saturated area.If try a clean importation of compress the nerve dollar after the adjustment get into flat area, make its exportation withdraw the saturated area of transformation function, can change the shape of error margin function, make the adjustment escape from flat area thus.The concrete way of doing which carries out this way of thinking BE, lead to go into an abrupt degree factor in the original transformation function.The picture identifies process to include:The picture prepares a processing, picture to partition,The picture characteristic sample, picture classification.The BP network structure has 3 F:Importation layer, implicit layer, exportation layer, such as figure 6-12 show.Withdrew to the handwritten numeral 5×5= be 25 characteristics, through level after cast shadow only 5 characteristic values be the importation of nerve network, so importation node is 5, according to implicit layer piece about input node 200% relation, the implicit layer takes 10 nodes, output a layer to take 4 nodes, this 4 exportations is four binary system number, representative the output's numerical type of the nerve network.Currently this network, study velocity because of improving BP calculate way very quick, is very high to be suffused with the ability of turn. |
PDF Full Text Request