Kwemakore, kudzidza kwakadzama kwave kuita misoro mu tech. Uye, zviri nyore kunzwisisa kuti sei.
Iri bazi rehungwaru hwekugadzira riri kushandura zvikamu kubva kune zvehutano kusvika kumabhanga kusvika kune zvekufambisa, zvichiita kufambira mberi kwaisambofungwa.
Kudzidza kwakadzama kwakavakirwa paseti yeakaomesesa algorithms anodzidza kuburitsa uye kufanotaura akaomesesa mapatani kubva kune yakakura mavhoriyamu edata.
Tichatarisa akanakisa gumi neshanu akadzika ekudzidza algorithms mune ino post, kubva kuConvolutional Neural Networks kuenda kuGenerative Adversarial Networks kune Yakareba Yenguva Yenguva Memory network.
Iyi positi ichapa ruzivo rwakakosha kuti iwe uri a anotanga kana nyanzvi yekudzidza kwakadzama.
1. Transformer Networks
Transformer network yachinja computer vision uye mashandisirwo emutauro wechisikigo (NLP). Vanoongorora data rinouya uye vanoshandisa maitiro ekutarisa kutora hukama hurefu. Izvi zvinovaita kuti vakurumidze kupfuura zvakajairika kutevedzana-kune-kutevedzana modhi.
Transformer network yakatanga kutsanangurwa mubhuku rinonzi "Attention Is All You need" naVaswani et al.
Iwo anosanganisira encoder uye decoder (2017). Iyo transformer modhi yakaratidza kuita mune dzakasiyana siyana dzeNLP application, kusanganisira manzwiro ongororo, kuiswa kwemavara, uye kushandura nemuchina.
Transformer-based modhi inogona zvakare kushandiswa mukuona komputa kune maapplication. Vanogona kuita kucherechedzwa kwechinhu uye kutora mifananidzo.
2. Yakareba Yenguva Ipfupi Memory Networks (LSTMs)
Yakareba Yenguva Yenguva Memory Networks (LSTMs) imhando ye neural network kunyanya kuvakwa kubata sequential input. Ivo vanonzi "nguva pfupi pfupi" nekuti vanogona kurangarira ruzivo kubva kare vachikanganwawo ruzivo rusina basa.
LSTMs inoshanda kuburikidza nemamwe "magedhi" anotonga kuyerera kweruzivo mukati metiweki. Zvichienderana nekuti ruzivo rwakakosha here kana kuti kwete, magedhi aya anogona kupinza kana kudzivirira.
Iyi nzira inogonesa LSTMs kuyeuka kana kukanganwa ruzivo kubva panguva yapfuura nhanho, iyo yakakosha kumabasa akaita sekuziva kutaura, kugadzirwa kwemutauro wechisikigo, uye nguva yekufanotaura.
LSTMs anobatsira zvakanyanya mune chero mamiriro ezvinhu apo iwe une sequential data inofanirwa kuongororwa kana kufanotaura. Iwo anowanzo shandiswa muzwi rekuziva software kushandura mazwi anotaurwa kuita zvinyorwa, kana mukati pamusika wemari ongororo yekufanotaura mitengo yeramangwana zvichibva pane yapfuura data.
3. Kuzvironga Mepu (SOMs)
MaSOM imhando yekugadzira neural network inogona kudzidza uye inomiririra data yakaoma munzvimbo yakaderera-dimensional. Nzira yacho inoshanda nekushandura data rekuisa repamusoro-soro kuita gidhi rine mativi maviri, neyuniti imwe neimwe kana neuron inomiririra chikamu chakasiyana chenzvimbo yekuisa.
Maneuroni akabatanidzwa pamwe chete uye anogadzira chimiro chepamusoro, achivabvumira kudzidza uye kugadzirisa kune data rekuisa. Saka, SOM yakavakirwa pakudzidza kusingatarisirwe.
Iyo algorithm haidi yakanyorwa data kudzidza kubva. Pane kudaro, inoshandisa zviverengero zve data rekuisa kuwana mapatani uye kuwirirana pakati pezvakasiyana.
Munguva yekudzidziswa nhanho, neurons inokwikwidza kuve yakanakisa chiratidzo cheiyo data yekuisa. Uye, ivo vanozvironga voga kuva chimiro chine musoro. MaSOM ane huwandu hwakawanda hwekushandisa, kusanganisira mufananidzo uye kuzivikanwa kwekutaura, kuchera data, uye kucherechedzwa kwemaitiro.
Vanobatsira kuona data yakaoma, kuunganidza zvine hukama data mapoinzi, uye kuona zvisirizvo kana kunze.
4. Deep Reinforcement Learning
Deep Kusimbisa Kudzidza imhando yekudzidza muchina umo mumiririri anodzidziswa kuita sarudzo zvichienderana nemubairo system. Inoshanda nekuita kuti mumiriri adyidzane nenzvimbo dzayo uye kudzidza kuburikidza nekuyedza uye kukanganisa.
Mumiririri anopihwa mubairo pane chese chaanoita, uye chinangwa chayo ndechekudzidza nzira yekuvandudza mabhenefiti ayo nekufamba kwenguva. Izvi zvinogona kushandiswa kudzidzisa vamiririri kutamba mitambo, kutyaira mota, uye kunyange kubata marobhoti.
Q-Kudzidza inzira inozivikanwa yeDeep Reinforcement Learning. Inoshanda nekuongorora kukosha kwekuita chimwe chiitiko mune imwe nyika uye kugadzirisa iyo fungidziro sezvo mumiriri anodyidzana nenharaunda.
Mumiririri anobva ashandisa fungidziro idzi kuona kuti ndechipi chiito chingangounza mubairo mukuru. Q-Kudzidza yakashandiswa kudzidzisa vamiririri kutamba mitambo yeAtari, pamwe nekuvandudza kushandiswa kwesimba munzvimbo dze data.
Deep Q-Networks imwe yakakurumbira Deep Reinforcement Learning nzira (DQN). DQNs dzakafanana neQ-Kudzidza mukuti vanofungidzira maitiro ekuita vachishandisa yakadzika neural network pane tafura.
Izvi zvinovagonesa kubata neakakura, akaomerwa marongero ane akawanda mamwe maitiro. DQNs dzakashandiswa kudzidzisa vamiririri kutamba mitambo yakadai seGo uye Dota 2, pamwe nekugadzira marobhoti anogona kudzidza kufamba.
5. Recurrent Neural Networks (RNNs)
RNNs imhando yeneural network inogona kugadzirisa sequential data uchichengeta mamiriro emukati. Kurangarira kwakafanana nomunhu anorava bhuku, apo shoko rimwe nerimwe rinogayiwa mukuwirirana naro rakauya pamberi paro.
Naizvozvo maRNN akanakira mabasa akaita sekuziva kutaura, kududzira mutauro, uye kunyange kufanotaura izwi rinotevera mumutsara.
MaRNN anoshanda nekushandisa mhinduro zvishwe kuti abatanidze kuburitsa kwega kwega nhanho kudzokera kune yekuisa yeinotevera nguva nhanho. Izvi zvinoita kuti network ishandise ruzivo rwekutanga nhanho yekuzivisa fungidziro yayo yematanho enguva yemberi. Nehurombo, izvi zvinoreva zvakare kuti maRNN ari panjodzi yekunyangarika gradient nyaya, umo ma gradients anoshandiswa pakudzidziswa anova madiki uye network inonetsekana kudzidza hukama hwenguva refu.
Kunyangwe ichi chimanikidziro chiri pachena, maRNN akawana kushandiswa mune dzakasiyana siyana dzekushandisa. Aya maapplication anosanganisira kugadzirwa kwemutauro wechisikigo, kuzivikanwa kwekutaura, uye kunyangwe kugadzirwa kwemimhanzi.
Shandurudzo google, semuenzaniso, inoshandisa RNN-based system kushandura mitauro yese, nepo Siri, mubatsiri chaiye, anoshandisa RNN-based system kuona inzwi. MaRNN akashandiswawo kufanotaura mitengo yemasheya uye kugadzira zvinyorwa zvechokwadi uye magiraidhi.
6. Capsule Networks
Capsule Networks rudzi rutsva rweneural network dhizaini inogona kuona mapatani uye kuwirirana mune data zvakanyanya. Vanoronga neurons kuita "capsules" inoisa mamwe maficha ekuisa.
Nenzira iyi vanogona kufanotaura chokwadi. Capsule Networks inobvisa zvishoma nezvishoma zvakaomarara zvivakwa kubva kudhata rekuisa nekushandisa akawanda akaturikidzana emakapisi.
Capsule Networks 'maitiro anovagonesa kudzidza hierarchical inomiririra yezvakapihwa. Ivo vanogona kurodha zvakafanira hukama hwepakati pakati pezvinhu zviri mukati memufananidzo nekutaurirana pakati pemakapisi.
Kuzivikanwa kwechinhu, kupatsanurwa kwemifananidzo, uye kugadzirwa kwemutauro wechisikigo zvese zvinoshandiswa zveCapsule Networks.
Capsule Networks ine mukana wekushandiswa mukati kuzvidzivirira kutyaira michina. Ivo vanobatsira sisitimu mukuziva nekusiyanisa pakati pezvinhu zvakaita semota, vanhu, uye zviratidzo zvemumigwagwa. Aya masisitimu anogona kudzivirira kubonderana nekuita fungidziro dzakanyatsojeka nezvemaitiro ezvinhu zviri munzvimbo yazvo.
7. Variational Autoencoders (VAEs)
MaVAE imhando yemudziyo wakadzama wekushandisa unoshandiswa pakudzidza usina mutariri. Nekukodha data munzvimbo yakaderera-dimensional uye vozoidzoreredza muchimiro chepakutanga, vanogona kudzidza kuona mapatani mudhata.
Vakafanana nen’anga inogona kushandura tsuro kuva ngowani yozodzokera kuva bhuru! MaVAE anobatsira pakugadzira zviono zvinoonekwa kana mimhanzi. Uye, ivo vanogona kushandiswa kugadzira data nyowani inofananidzwa neyekutanga data.
MaVAE akafanana neakavanzika codebreaker. Vanogona kuziva zviri pazasi chimiro che data nekuchityora-tyora kuita zvidimbu zvakapfava, sekuti dambanemazwi rinoputswa sei. Vanogona kushandisa ruzivo irworwo kuvaka data nyowani inoita seyekutanga mushure mekunge varongedza zvikamu.
Izvi zvinogona kubatsira pakumanikidza mafaera akakura kana kugadzira mifananidzo mitsva kana mimhanzi mune imwe chimiro. MaVAE anogona zvakare kugadzira zvinyowani zvemukati, senge nhau dzenhau kana nziyo dzemimhanzi.
8. Generative Adversarial Networks (GANs)
MaGAN (Generative Adversarial Networks) imhando yehurongwa hwekudzidza hwakadzama hunogadzira data idzva rakafanana nerekutanga. Ivo vanoshanda nekudzidzisa ma network maviri: jenareta uye network yekusarura.
Iyo jenareta inogadzira data nyowani inofananidzwa neyekutanga.
Uye, musarura anoedza kusiyanisa pakati pekutanga uye data yakagadzirwa. Iwo ma network maviri akadzidziswa tandem, nejenareta ichiedza kunyengedza musarura uye musarura achiedza kunyatsoziva data rekutanga.
Funga maGAN semuchinjiko pakati peforoji nemutikitivha. Jenareta inoshanda zvakafanana nefoja, ichigadzira mifananidzo mitsva yakafanana neyepakutanga.
Anosarura anoita semutikitivha, achiedza kusiyanisa pakati pemifananidzo yechokwadi nekunyepedzera. Manetiweki maviri aya akadzidziswa tandem, jenareta ichinatsiridza pakugadzira manyepo anonzwisisika uye musarura anovandudza pakuvaziva.
MaGAN ane mashandisiro akati wandei, kubva pakugadzira mifananidzo yechokwadi yevanhu kana mhuka kusvika pakugadzira mimhanzi mitsva kana kunyora. Iwo anogona zvakare kushandiswa kuwedzera data, izvo zvinosanganisira kubatanidza inogadzirwa data neiyo chaiyo data kuti ivake yakakura dhatabheti yekudzidzira modhi yekudzidzira.
9. Yakadzika Q-Networks (DQNs)
Yakadzika Q-Networks (DQNs) imhando yekuita sarudzo yekusimbisa kudzidza algorithm. Ivo vanoshanda nekudzidza Q-basa rinofanotaura mubairo unotarisirwa kuita chimwe chiitiko mune imwe mamiriro.
Iyo Q-basa inodzidziswa nekuyedza uye kukanganisa, neiyo algorithm ichiedza zviito zvakasiyana uye kudzidza kubva pane zvabuda.
Zvione sea vhidhiyo yemutambo hunhu hunoyedza nezviito zvakasiyana uye kuwana kuti ndezvipi zvinotungamira mukubudirira! DQNs inodzidzisa Q-basa vachishandisa yakadzika neural network, ichivaita maturusi anoshanda kune yakaoma kuita sarudzo mabasa.
Vakatokunda shasha dzevanhu mumitambo yakaita seGo and chess, pamwe nemumarobhoti nemotokari dzinozvityaira. Saka, mune zvese, maDQN anoshanda nekudzidza kubva kune ruzivo kuti vawedzere hunyanzvi hwavo hwekuita sarudzo nekufamba kwenguva.
10. Radial Basis Function Networks (RBFNs)
Radial Basis Function Networks (RBFNs) imhando yeneural network iyo inoshandiswa kufungidzira mabasa uye kuita mabasa ekuisa. Ivo vanoshanda nekushandura iyo data yekupinza kuita yakakwirira-dimensional nzvimbo vachishandisa muunganidzwa we radial base mabasa.
Kubuda kwenetiweki musanganiswa wemutsara webasa rekutanga, uye yega yega radial base basa inomiririra nzvimbo yepakati munzvimbo yekupinda.
MaRBFN anonyanya kushanda kumamiriro ezvinhu ane kudyidzana kwakaoma kwekuisa-zvabuda, uye anogona kudzidziswa pachishandiswa nzira dzakasiyana siyana, kusanganisira kudzidza kunotariswa uye kusingatarisirwe. Vave vachishandiswa chero chinhu kubva kukufungidzira kwezvemari kusvika kumufananidzo uye kuzivikanwa kwekutaura kune kuongororwa kwekurapa.
Funga nezveRBFN seGPS system inoshandisa nhevedzano yemapoinzi eanchor kuwana nzira yayo munzvimbo dzakaoma. Kubuda kwetiweki kusanganiswa kweanchor point, inomira mukati meiyo radial base mabasa.
Isu tinokwanisa kuongorora kuburikidza neruzivo rwakaoma uye kugadzira fungidziro chaiyo yekuti mamiriro achafamba sei nekushandisa maRBFN.
11. Multilayer Perceptrons (MLPs)
Iyo yakajairika fomu yeneural network inodaidzwa kuti multilayer perceptron (MLP) inoshandiswa kune inotariswa mabasa ekudzidza senge kurongedza uye kudzoreredza. Ivo vanoshanda nekurongedza akati wandei matinji emanodhi akabatana, kana neurons, nechero imwe neimwe isina mutsara ichichinja data inouya.
MuMLP, neuron yega yega inowana mapindiro kubva mutsinga dziri pazasi uye inotumira chiratidzo kune neuroni dziri pamusoro. Imwe neimwe inobuda neuron inotemerwa uchishandisa activation basa, iyo inopa iyo network isina mutsara.
Ivo vanokwanisa kudzidza kumiririra kwakaomarara kweiyo data yekupinza sezvo vachigona kuve neakati wandei akavanzwa akaturikidzana.
MLPs dzakashandiswa kumabasa akasiyana-siyana, akadai sekuongorora manzwiro, kuonekwa kwehutsotsi, uye kuzivikanwa kwezwi nemifananidzo. MLPs inogona kufananidzwa neboka revaongorori vari kushanda pamwechete kupaza nyaya yakaoma.
Pamwe chete, vanogona kubatanidza chokwadi uye kugadzirisa mhosva kunyangwe hazvo paine imwe nzvimbo yehunyanzvi.
12. Convolutional Neural Networks (CNNs)
Mifananidzo nemavhidhiyo zvinogadziriswa pachishandiswa convolutional neural network (CNNs), chimiro cheneural network. Dzinoshanda nekushandisa seti yemafirita anogona kudzidziswa, kana kernels, kutora hunhu hwakakosha kubva kune data rekuisa.
Masefa anotsvedza pamusoro pemufananidzo wekuisa, achiita convolutions kugadzira mepu inobata zvakakosha zvemufananidzo.
Sezvo maCNN achikwanisa kudzidza mamiririro emhando yemifananidzo, anonyanya kubatsira kumamiriro ezvinhu anosanganisira hukuru hwe data yekuona. Zvishandiso zvinoverengeka zvakazvishandisa, sekuonekwa kwechinhu, kukamurwa kwemifananidzo, uye kuona chiso.
Funga nezveCNNs semupendi anoshandisa mabhurashi akati wandei kugadzira hunyanzvi. Bhurashi rega rega ikernel, uye muimbi anogona kuvaka mufananidzo wakaoma, unogoneka nekusanganisa mhodzi dzakawanda. Isu tinokwanisa kutora hunhu hwakakosha kubva kumifananidzo uye nekuishandisa kufanotaura nemazvo zviri mukati memufananidzo nekushandisa CNNs.
13. Deep Belief Networks (DBNs)
MaDBN imhando yeneural network iyo inoshandiswa kune isina kutariswa mabasa ekudzidza akadai sekuderedza dimensionality uye chimiro chekudzidza. Iwo anoshanda nekurongedza akati wandei akaturikidzana eRestricted Boltzmann Machines (RBMs), ari maviri-layer neural network inokwanisa kudzidza kudzoreredza data rekuisa.
MaDBN anobatsira zvakanyanya kune akakwira-dimensional data nyaya nekuti anogona kudzidza compact uye inonyatsomiririra inomiririra yekuisa. Akashandiswa kune chero chinhu kubva pakuzivikanwa kwezwi kusvika pakurongwa kwemifananidzo kusvika pakuwanikwa kwezvinodhaka.
Semuenzaniso, vatsvakurudzi vakashandisa DBN kuti vafungidzire hukama hunosungirirwa hwevanoda mishonga kune estrogen receptor. Iyo DBN yakadzidziswa muunganidzwa wemakemikari maitiro uye anosunga affinities, uye yakakwanisa kufanotaura nemazvo hukama hunosunga hwevanoda zvinodhaka.
Izvi zvinoratidzira kushandiswa kweDBN mukuvandudza zvinodhaka uye mamwe epamusoro-dimensional data application.
14. Autoencoders
Autoencoders maneural network ayo anoshandiswa kune asina kutariswa ekudzidza mabasa. Ivo vanoitirwa kuvakazve data rekuisa, izvo zvinoreva kuti ivo vanozodzidza kukodha iyo ruzivo mune inomiririra compact uye vozoidzoreredza mune yekutanga kuisa.
Autoencoders inoshanda zvakanyanya pakudzvanya data, kubvisa ruzha, uye kuona zvisina kunaka. Iwo anogona zvakare kushandiswa pakudzidza maficha, uko autoencoder's compact inomiririra inopihwa mune inotariswa basa rekudzidza.
Funga ma autoencoders kuve vadzidzi vanotora manotsi mukirasi. Mudzidzi anoteerera hurukuro uye anonyora pasi pfungwa dzinonyanya kukosha nenzira pfupi uye inoshanda.
Gare gare, mudzidzi angadzidza ndokuyeuka chidzidzo chacho achishandisa manotsi avo. Iyo autoencoder, kune rumwe rutivi, inoisa iyo data yekupinda mune inomiririra compact iyo inogona kuzoshandiswa kune zvakasiyana zvinangwa sekuona kusinganzwisisike kana kudzvanya data.
15. Yakarambidzwa Boltzmann Machines (RBMs)
RBMs (Restricted Boltzmann Machines) imhando yekugadzira neural network iyo inoshandiswa kumabasa asina kutarisirwa ekudzidza. Iwo anoumbwa nedhizaini inooneka uye yakavanzika layer, ine neuroni mune yega yega, yakabatana asi isiri mukati meiyo imwechete.
MaRBM anodzidziswa achishandisa nzira inozivikanwa sekusiyana-siyana, iyo inosanganisira kushandura huremu pakati pezvinooneka uye zvakavanzwa zvidimbu kuitira kukwidziridza mukana weiyo data yekudzidziswa. MaRBM anogona kugadzira data nyowani mushure mekudzidziswa nesampling kubva mukugovera kwakadzidzwa.
Kuzivikanwa kwemufananidzo nematauriro, kusefa kwekudyidzana, uye kuona zvisinganzwisisike zvese zvinoshandiswa zvakashandisa maRBM. Iwo akashandiswawo mumasisitimu ekurudziro kugadzira kurudziro yakaenderana nekudzidza maitiro kubva mumaitiro emushandisi.
MaRBM akashandiswawo mukudzidza kwechimiro kugadzira compact uye inoshanda inomiririra yeakakwira-dimensional data.
Pedzisa-Uye uye Anovimbisa Zviitiko paHorizon
Nzira dzekudzidza dzakadzama, dzakadai seConvolutional Neural Networks (CNNs) uye Recurrent Neural Networks (RNNs), ndedzimwe dzenzira dzepamusoro-soro dzehungwaru. CNNs dzakashandura mapikicha uye kuzivikanwa kwekuteerera, nepo maRNN akafambira mberi zvakanyanya mumutauro wechisikigo kugadzirisa uye kutevedzana kwedata data.
Nhanho inotevera mukushanduka kwemaitiro aya ingangove yakatarisana nekuvandudza kugona kwavo uye scalability, ichivabvumira kuongorora makuru uye akaomarara dhatasets, pamwe nekusimudzira kududzira kwavo uye kugona kudzidza kubva kune isina kunyorwa data.
Kudzidza kwakadzama kune mukana wekubvumira kubudirira muminda yakadai sehutano, mari, uye kuzvitonga masisitimu sezvainofambira mberi.
Leave a Reply