Ụdị dabere na ike
Ụdị dabere na ume (EBMs) na-amụta ọrụ 'ike' nke na-ekenye ụkpụrụ dị ala na data bara uru yana ụkpụrụ dị elu na data na-enweghị atụ, na-akọwa nkesa ihe gbasara omume na-enweghị ịmanye ya ka ọ dị mfe ịhazi.
Nchịkọta
This flexibility makes them a unifying lens for much of machine learning, from classifiers to generative models.
Ime miri emi
Nlereanya dabere na ume na-akọwa ihe gbasara puru omume site na nkesa Boltzmann (Gibbs): p (x) dabara na exp (-E (x)), ebe E (x) bụ ọrụ ike mụtara, na-abụkarị netwọkụ akwara. Ọzụzụ na-akwada ike nke ezigbo data ma na-ebuli ike nke ihe ọ bụla ọzọ. Ihe ejidere bụ ọrụ nkebi Z, nchikota ma ọ bụ ngwakọta nke exp(-E(x)) karịa ntinye niile enwere ike, nke na-adịkarị mfe ịgbakọ. Ya mere, a na-azụ EBMs na nsoro: ọdịiche dị iche iche, nha nha nha, ma ọ bụ ntule mkpọtụ na-emegiderịta onwe ya, na atụpụta ya site na usoro MCMC dị ka Langevin dynamics na-esote ike gradient. Ihe atụ kpochapụwo gụnyere netwọk Hopfield na igwe Boltzmann amachibidoro; Ọrụ ọgbara ọhụrụ na-ejikọ EBM na ụdị mgbasa ozi, GAN, na ọbụna klas ndị nkịtị atụgharịgharịrị dị ka ọrụ ike.
Nghọta nka nka
Ihe nlereanya ahụ na-enye ihe gbasara puru omume p(x) = exp(-E(x))/Z. N'ihi na Z (ihe na-eme ka ihe ọ bụla na-eme ihe na-eme ka ọ bụrụ ihe na-eme ka ọ bụrụ ihe na-edozi ahụ) adịghị agwụ agwụ, ọ naghị esiri gị ike ịgbakọ ohere ozugbo. Kama, akara dakọtara na nlele Langevin na-erigbu na gradient nke log p(x) nhata -gradient nke E(x), yabụ Z ga-apụ. Langevin dynamics wee na-ewepụta ihe nlele site n'itinye ume ugboro ugboro na mgbada x na-agbakwunye mkpọtụ, na-aga na mpaghara ike dị ala, nke nwere ike dị elu.
Mmetụta atụmatụ
Ọnụ ego na mmefu ego
Mkpebi ihe owuwu ụlọ na-akwalite arụmọrụ yana ọnụ ahịa ọrụ ruo ọtụtụ afọ.
Mkpebi doro anya
Nkà mmụta nka na-enyere ndị otu egwuregwu aka ịhọrọ nchịkọta ziri ezi, ọ bụghị naanị nke kachasị ọhụrụ.
Quality akara
Nhọrọ injinia ka mma na-ebelata ihe omume ntụkwasị obi na mmepụta.
Ọdịnihu nke Ụdị Dabere Ike
Ndị EBM na-anụ ụtọ mmasị ọhụrụ n'ihi na ha na-enye àkwà mmiri dị n'etiti ụdị mgbasa ozi, ụdị mmepụta akara, na netwọk ịkpa ókè, akara nke ụdị mgbasa ozi na-amụta bụ n'ezie gradient ume. Na-atụ anya ka usoro ngwakọ ndị ọzọ na-eji ọrụ ike maka mgbanwe, mgbochi ndị nwere ike ịmegharị (ijikọta ọtụtụ ume na-eduzi ọgbọ), nlele ka mma na ngwa ngwa karịa MCMC, yana ngwa n'echiche na nhazi ebe 'chọta nhazi ike kachasị dị ala' na-egosipụta njikarịcha na afọ ojuju.
Mmejuputa n'ezie n'ụwa
Netwọk Hopfield na-arụ ọrụ dị ka ebe nchekwa na-akpakọrịta nke na-echeta usoro echekwara site na ntinye mkpọtụ ma ọ bụ nke akụkụ site na ịbanye n'ime ọnọdụ ike dị ala.
Igwe Boltzmann amachibidoro ejiri mee ihe n'akụkọ ihe mere eme maka nzacha ọnụ na ịmalite netwọọdụ nkwenye miri emi.
Ịsụgharịgharị ọkọlọtọ ọkọlọtọ dị ka ihe nlere dabere na ume (usoro JEM) iji melite nhazi, ike siri ike na nchọpụta nkesa adịghị.
Amụma ahaziri ahazi na afọ ojuju mmachi, ebe a na-achọta azịza site na ibelata ike mmụta n'ọtụtụ mgbanwe mmekọrịta (dịka ọmụmaatụ, nleba anya ma ọ bụ nhazi)
Ihe ize ndụ & okporo ụzọ nche
Ịkwalite otu akara ngosi nwere ike zoo adịghị ike sistemụ sara mbara.
A na-eledakarị ihe akụrụngwa na ụgwọ ọrụ anya.
Ọdịiche nchekwa na nleba anya nwere ike itolite ka sistemu na-adịwanye mgbagwoju anya.
Map mmejuputa
Kọwaa latency, ịdịmma na ebumnuche ọnụ ahịa tupu mmejuputa ya.
Benchmark n'okpuru ibu dị adị na ọnọdụ data.
Nleba anya akụrụngwa maka mperi, ịkpafu na mmetụta onye ọrụ.
Kwadebe ụzọ nzaghachi azụghachi azụ na ihe omume tupu ịchachaa.
Nọgide na-eme nchọpụta
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the Energy-Based Models quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ntuziaka na-esote
Ụdị Ọdịbendị dabere na akara
Ajụjụ a na-ajụkarị
What is Energy-Based Models?
Ụdị dabere na ume (EBMs) na-amụta ọrụ 'ike' nke na-ekenye ụkpụrụ dị ala na data bara uru yana ụkpụrụ dị elu na data na-enweghị atụ, na-akọwa nkesa ihe gbasara omume na-enweghị ịmanye ya ka ọ dị mfe ịhazi. Mgbanwe a na-eme ka ha bụrụ oghere na-ejikọta ọnụ maka ọtụtụ mmụta igwe, site na nhazi ọkwa ruo na ụdị mmepụta.
N'ihe atụ nke ike dabere, kedu ka ike si metụta ihe gbasara ike nke ebe data?
Site na nkesa Boltzmann, p (x) dabara na exp(-E (x)), yabụ na-ekenye data ezi uche dị na ya ike dị ala na ohere dị elu.
Gịnị na-eme ka ọzụzụ ike dabere ụdị siri ike?
Ịgbakọ Z chọrọ nchikota ma ọ bụ ijikọta exp(-E(x)) n'ofe ntinye niile, nke na-adịkarịghị ekwe omume, na-amanye usoro ọzụzụ dị ka.
Kedu usoro nlere anya a na-ejikarị wepụta sample site na EBM?
Langevin dynamics ugboro ugboro na-akpali samples gbadata na ike gradient ka ọ na-agbakwunye mkpọtụ, na-agbakọta na mpaghara ndị nwere obere ume.
Kedu ihe kpatara usoro dị ka nha nhata nwere ike isi zere ịgbakọ ọrụ nkebi Z?
Ebe Z adabereghị na x, iche log p(x) n'ihe gbasara x na-akagbu ya, na-ahapụ naanị ike gradient, bụ traktị.
Kedu n'ime ihe ndị a bụ ihe atụ nke dabere na ume?
Netwọk Hopfield bụ ihe nlere mbụ dabere na ume nke na-echekwa ụkpụrụ dị ka steeti ike dị ala wee cheta ha site na ibelata ume.