Online ati Lile Negetifu Mining
Iwakusa odi lile mu alaye ti o pọ julọ, awọn apẹẹrẹ ti o nira-lati ṣe iyatọ lati ṣe ikẹkọ lori dipo jafara akitiyan lori awọn ti o rọrun awoṣe naa ti ni ẹtọ tẹlẹ.
Akopọ
O jẹ ẹtan ti o mu ki ẹkọ metric ati wiwa nkan papọ ni kiakia ati deede.
Jin Dive
Nigbati ikẹkọ pẹlu meteta tabi awọn adanu iyatọ, pupọ julọ awọn odi ti a ṣe ayẹwo laileto ti jinna si oran, nitorinaa wọn gbejade pipadanu odo ko si si gradient, awọn iduro ikẹkọ. Iwakusa ti ko dara ṣe atunṣe eyi nipa yiyan awọn odi lile: awọn apẹẹrẹ ti o jẹ aṣiṣe ti o sunmọ isunmọ. Ni iwakusa aisinipo, o ṣayẹwo lorekore dataset lati wa iwọnyi, eyiti o lọra ti o si lọ. Iwakusa ori ayelujara ṣe iṣiro wọn lori fifo laarin ipele kekere kọọkan: lẹhin ti o kọja siwaju, o wo gbogbo awọn ijinna meji-meji ni ipele ki o yan awọn irufin ti o nira julọ. FaceNet ṣafihan iwakusa ologbele-lile, yiyan awọn odi ti o jinna ju rere lọ ṣugbọn sibẹ inu ala, yago fun aisedeede ti awọn odi lile lile le fa ni kutukutu ikẹkọ.
Imọ-imọ-ẹrọ
Online mining exploits the batch you already computed. With B embeddings you get a B-by-B distance matrix essentially for free, so you can evaluate huge numbers of candidate triplets per step. Batch-hard mining selects, for each anchor, the farthest positive and the nearest negative in the batch. Iwakusa ologbele-lile dipo awọn idiwọ odi lati dubulẹ laarin ijinna rere ati ijinna rere pẹlu ala, ti n ṣe agbejade aisi-odo ṣugbọn awọn gradients iduroṣinṣin. Larger batches give a richer pool of hard candidates, which is why batch size strongly affects metric-learning quality.
Ipa Ilana
Iye owo ati isuna
Awọn ipinnu faaji ṣe awakọ iṣẹ ati idiyele iṣẹ fun awọn ọdun.
Awọn ipinnu diẹ sii
Ẹkọ imọ-ẹrọ ṣe iranlọwọ fun awọn ẹgbẹ lati yan akopọ to tọ, kii ṣe ọkan tuntun nikan.
Iṣakoso didara
Awọn yiyan imọ-ẹrọ to dara julọ dinku awọn iṣẹlẹ igbẹkẹle ni iṣelọpọ.
Ojo iwaju ti Online ati Lile Negetifu Mining
Ilana naa, ikẹkọ lori ohun ti o ṣoro, bayi n ṣe ikẹkọ ikẹkọ ti ara ẹni ti o yatọ, nibiti awọn adagun nla inu-ipele (ati awọn banki iranti bi MoCo) pese awọn afiwera ti o nira laisi awọn aami. Awọn oniwadi n ṣatunṣe bi odi ṣe yẹ ki o le, niwọn igba ti awọn odi lile ju nigbagbogbo yipada lati jẹ ami ti ko tọ tabi awọn ohun rere ti o sunmọ-ẹda ti o ba ikẹkọ jẹ. Expect smarter, uncertainty-aware mining and synthetic hard negatives generated by the model itself, plus tighter integration with retrieval systems that mine hard negatives from real user queries.
Real-World imuse
Idanileko idanimọ oju: FaceNet nlo iwakusa ori ayelujara ologbele-lile lati kọ ẹkọ awọn ifibọ ti o ya awọn ẹni-kọọkan ti o jọra.
Wiwa nkan: SSD ati awọn aṣawari ti o jọra lo iwakusa odi lile lati dọgbadọgba ikun omi ti awọn apoti isale irọrun lodi si awọn apoti ohun toje.
Igbapada aye ipon: wiwa ati awọn ọna ṣiṣe RAG mi awọn iwe aṣẹ odi lile ti o jọmọ ṣugbọn kii ṣe, ti n mu olugba pada.
Awọn ọna ṣiṣe iṣeduro: awọn awoṣe awọn ohun elo mi ti olumulo ko tẹ ṣugbọn ti o jọra awọn nkan ti o tẹ, nkọ awọn iyatọ to dara julọ ni itọwo.
Awọn ewu & Awọn ọna iṣọ
Ṣiṣepe ala-ilẹ kan le tọju awọn ailagbara eto ti o gbooro.
Awọn ohun elo amayederun ati awọn idiyele itọju nigbagbogbo ni aibikita.
Aabo ati awọn ela akiyesi le dagba bi awọn eto ṣe di eka sii.
Ilana Ilana imuse
Ṣetumo lairi, didara, ati awọn ibi-afẹde idiyele ṣaaju imuse.
Aṣepari labẹ ẹru ojulowo ati awọn ipo data.
Abojuto ohun elo fun awọn aṣiṣe, fiseete, ati ipa olumulo.
Mura ipadasẹhin pada ati awọn ipa ọna esi iṣẹlẹ ṣaaju iwọn.
Tesiwaju Ṣiṣawari
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 Online and Hard Negative Mining 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
Itọsọna atẹle
Online ati Aisinipo Ẹya Nṣiṣẹ Skew
Awọn ibeere ti a beere nigbagbogbo
Kini Online ati Hard Negative Mining?
Iwakusa odi lile mu alaye ti o pọ julọ, awọn apẹẹrẹ ti o nira-lati ṣe iyatọ lati ṣe ikẹkọ lori dipo jafara akitiyan lori awọn ti o rọrun awoṣe naa ti ni ẹtọ tẹlẹ. O jẹ ẹtan ti o jẹ ki ẹkọ metiriki ati wiwa nkan ṣe apejọ ni iyara ati ni deede.
Kilode ti ọpọlọpọ awọn odi ti a ṣe ayẹwo laileto pese ifihan agbara ikẹkọ kekere ni pipadanu mẹta?
Awọn odi ti o rọrun joko daradara ni ikọja ala, ni itẹlọrun pipadanu tẹlẹ, ati pe o ṣe alabapin ni pataki ko si gradient, nitorinaa awọn iduro ikẹkọ.
Kini o ṣe iyatọ iwakusa ori ayelujara lati iwakusa aisinipo?
Iwakusa ori ayelujara ṣe iṣiro awọn ijinna meji-meji laarin ipele lọwọlọwọ ni igbesẹ kọọkan, lakoko ti iwakusa aisinipo lorekore ṣayẹwo gbogbo dataset ati pe o le di asan.
Kini odi 'ologbele-lile' bi a ti ṣalaye ni FaceNet?
Awọn aibikita ologbele-lile jinna si oran ju rere sibẹsibẹ ṣubu sinu ala, fifun iwulo, awọn gradients iduroṣinṣin laisi aisedeede ti awọn ọran ti o nira julọ.
Ni iwakusa ipele-lile, odi wo ni a yan fun oran kọọkan?
Iwakusa ipele-lile mu awọn ọran ti o nira julọ: fun oran kọọkan o gba rere ti o jinna julọ ati odi ti o sunmọ (julọ airoju) odi ninu ipele naa.
Kini idi ti iwọn ipele ti o tobi julọ ṣọ lati mu ilọsiwaju awọn abajade ikẹkọ metiriki mined?
Ipele nla kan n ṣe agbejade matrix ijinna meji-meji nla, nitorinaa awọn odi oludije diẹ sii wa si mi ti o nira julọ lati igbesẹ kọọkan.