Ozi na ebumnuche SimCLR
InfoNCE bụ mfu dị iche iche na-akuziri ihe atụ ijikọkọta ụzọ abụọ na-adakọ ọnụ na ịkwanye ụzọ abụọ ejikọtaghị ọnụ na ntinye oghere.
Nchịkọta
SimCLR is a landmark framework that used this loss to learn powerful image representations from unlabeled data, rivaling supervised pretraining.
Ime miri emi
InfoNCE (Mkpọtụ-Contrastive Estimation maka ozi gbasara ibe) na-azụ ihe ngbanwe ka ajụjụ na ezigbo mma ya nwee akara myirịta dị elu karịa ajụjụ a na ọtụtụ adịghị mma. Ọ bụ n'ezie softmax cross-entropy n'elu nha nha: maka arịlịka, ihe dị mma kwesịrị imeri megide ihe ọjọọ. SimCLR (2020) arụ ọrụ nke a maka onyonyo: were otu onyonyo, tinye mgbakwunye abụọ na-enweghị usoro iji mepụta ụzọ dị mma, gbaa ha abụọ site na ntinye koodu na-ekekọrịta yana isi amụma, wee jiri cross-entropy na-agbanwe agbanwe (NT-Xent, ụdị InfoNCE) yabụ echiche abụọ ahụ agbagoro na-adọta ebe onyonyo ndị ọzọ dị na batch na-eme dị ka adịghị mma. SimCLR gosipụtara na mmụba data siri ike, isi ntule na-adịghị ahụ anya, nnukwu ogbe nha, yana ọnọdụ okpomọkụ jikọtara ọnụ ka ụdị ndị na-ahụ maka onwe ha dakọtara na ndị a na-elekọta na ImageNet - na-enweghị akara ọ bụla n'oge ọzụzụ ọzụzụ.
Nghọta nka nka
NT-Xent na-agbakọ myirịta cosine n'etiti ntinye L2-normalized, na-ekewa site na okpomọkụ τ, ma tinye softmax cross-entropy na-emeso ihe dị mma dị ka klas ziri ezi n'etiti ihe atụ niile dị n'ime. Lower τ na-amụba nkesa na-ata ahụhụ karịa ihe ọjọọ. A na-eji isi ntule SimCLR (MLP) naanị n'oge ọzụzụ ma tụfuo ya ma emesịa - ihe nnọchianya tupu isi ebufe nke ọma. Nnukwu batches dị mkpa n'ihi na ha na-enye ọtụtụ ihe ọjọọ n'otu nzọụkwụ.
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 Ozi na ebumnuche SimCLR
Ebumnobi ndị na-emegiderịta onwe ha gbasapụrụ karịa SimCLR: CLIP na-ejikọta onyonyo na ederede site na iji InfoNCE n'ofe usoro, na otu mfu ahụ na-akpali ọdịyo, vidiyo, na ụdị iweghachite. Nnyocha ugbu a na-ebelata ịdabere na nnukwu batches na ọtụtụ ihe ọjọọ site na ebe nchekwa nchekwa (MoCo), ma ọ bụ wepụ ihe ọjọọ kpamkpam (BYOL, SimSiam, DINO). Na-atụ anya ka na-aga n'ihu na-agwakọta nke ọdịiche, ntugharị, na ihe nkpuchi pretraining, yana nhazi multimodal (ederede, onyonyo, ọdịyo) dịka oke oke maka ụdị ntọala.
Mmejuputa n'ezie n'ụwa
SimCLR na-ebu ụzọ zụọ ihe ngbanwe onyonyo na foto enweghị akara, wee na-emegharị nke ọma na obere ihe akpọrọ maka nhazi.
CLIP na-eji ebumnobi Ozi iji dakọọ onyonyo na nkebi okwu ha, na-eme ka nhazi onyonyo enweghị atụ.
Iwulite nchọta/nweghachi anya ebe ihe onyonyo ndị yiri ya nọdụrụ ọnụ n'ime oghere mmụta mmụta.
Nlekọta onwe onye na-elekọta maka ahụike ahụike ma ọ bụ onyonyo satịlaịtị ebe akara dị ụkọ mana data adịchaghị ukwuu.
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 InfoNCE and SimCLR Objectives 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
Lookhead na Lion Optimizers
Ajụjụ a na-ajụkarị
What is InfoNCE and SimCLR Objectives?
InfoNCE bụ mfu dị iche iche na-akuziri ihe atụ ijikọkọta ụzọ abụọ na-adakọ ọnụ na ịkwanye ụzọ abụọ ejikọtaghị ọnụ na ntinye oghere. SimCLR bụ akara ama ama ejirila mfu a mụta ihe onyonyo siri ike site na data enweghị akara, na-amata ọzụzụ izizi a na-elekọta.
Kedu ihe ebumnobi InfoNCE na-akwado ihe nlereanya ime?
InfoNCE bụ softmax n'elu myirịta nke na-akwụ ụgwọ myirịta dị elu maka ezigbo myirịta dị mma na nke dị ala maka adịghị mma.
Na SimCLR, kedu ka esi emepụta ụzọ dị mma?
SimCLR na-ewepụta ụzọ dị mma site na echiche abụọ agbagoro nke otu ihe onyonyo; ihe oyiyi ndị ọzọ na-eje ozi dị ka ndị na-adịghị mma.
Kedu ọrụ okpomọkụ τ na-arụ na NT-Xent / Ozi?
Ịkekọrịta myirịta site na τ tupu softmax na-achịkwa ka ọnwụ ahụ si akọwapụta ihe dị mma na ihe ọjọọ siri ike.
Kedu ihe ga-eme isi ntule SimCLR ka emechara ọzụzụ?
SimCLR chọpụtara na njirimara ahụ tupu ebufe isi ntule ka mma, yabụ a na-atụfu isi ya mgbe emechara ọzụzụ.
Kedu ihe kpatara SimCLR ji dabere na nnukwu batch?
Na SimCLR ihe onyonyo ndị ọzọ dị na batch na-eme ihe na-adịghị mma, yabụ batches buru ibu na-enyekwu ihe ọjọọ na mmụta siri ike.