Iṣeeṣe odiwọn
Isọdiwọn tumọ si awọn iṣeeṣe ti a sọ ti awoṣe baramu otitọ: nigbati o sọ 70%, iṣẹlẹ yẹ ki o ṣẹlẹ nipa 70% ti akoko naa.
Akopọ
It matters because accurate confidence drives good decisions in medicine, finance, and risk-sensitive AI.
Jin Dive
Awoṣe le jẹ deede sibẹsibẹ ti koṣe calibrated. Awọn nẹtiwọọki ti o jinlẹ ode oni jẹ olokiki fun igbẹkẹle apọju, ti njade 99% lori awọn asọtẹlẹ ti o jẹ ẹtọ ti o kere pupọ nigbagbogbo. Iṣatunṣe iwọntunwọnsi eyi nipasẹ awọn asọtẹlẹ bucketing nipasẹ igbẹkẹle ati ṣayẹwo igbohunsafẹfẹ ti a ṣe akiyesi ni garawa kọọkan. Awọn igbero aworan igbẹkẹle ti a sọtẹlẹ si gangan; awoṣe calibrated daradara kan joko lori akọ-rọsẹ. Aṣiṣe Iṣatunṣe Ireti (ECE) ṣe akopọ aafo naa gẹgẹbi aropin iwuwo kọja awọn apoti. Awọn atunṣe wa ni awọn adun meji: awọn ọna post-hoc bi Platt scaling (ti o baamu iyipada logistic), iwọn otutu iwọn otutu (pinpin awọn iṣiro nipasẹ scalar T ti o kọ ẹkọ), ati isotonic regression (igbesẹ monotonic kan ti o yẹ); ati awọn ọna ikẹkọ-akoko bi aami didan tabi awọn adanu igbelewọn to dara. Isọdiwọn ati deede jẹ awọn ibi-afẹde pato, ati ilọsiwaju ọkan ko nilo ilọsiwaju.
Imọ-imọ-ẹrọ
Pipin iwọn otutu jẹ ẹṣin iṣẹ fun awọn netiwọki nkankikan: pin awọn iwe-iṣaaju-softmax nipasẹ iwọn otutu T kan ti a kọ ẹkọ, lẹhinna tun-softmax. T> 1 jẹ ki awọn ipinpinpin igbẹkẹle rọra, T <1 mu wọn pọ. Ni pataki T jẹ ibamu lori data afọwọsi lati dinku iṣeeṣe log-odi ati pe ko yipada eyiti kilasi bori, nitorinaa deede ko fọwọkan lakoko ti awọn iṣeeṣe di ooto. Paramita ẹyọkan rẹ jẹ ki o jẹ data-daradara ati pe ko ṣee ṣe lati overfit.
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 Iṣeeṣe odiwọn
Bi AI ti nwọle awọn iyipo giga-giga, isọdọtun n gbe lati ero lẹhin si ibeere. Iṣẹ n pọ si ọna calibrated igbẹkẹle awoṣe ede ti o tobi ati aidaniloju ọrọ, isọdiwọn labẹ iyipada pinpin, ati isọdi ọgbọn-ẹgbẹ nitori awọn iṣeeṣe jẹ ododo ni gbogbo awọn agbejade. Reti awọn metiriki odiwọn lẹgbẹẹ deede ni awọn kaadi awoṣe ati awọn iṣayẹwo ilana, pẹlu isọpọ wiwọ pẹlu asọtẹlẹ ibamu ati asọtẹlẹ yiyan ki awọn eto le yago fun igbẹkẹle nigbati igbẹkẹle ododo wọn lọ silẹ.
Real-World imuse
Iṣẹ oju-ọjọ kan ṣe idaniloju pe asọtẹlẹ awọn ọjọ ni 30% ojo n rii ni gangan ojo nipa 30% ti akoko naa, ibi-afẹde isọdọtun iwe-ẹkọ.
Awoṣe aiyipada-kirẹditi jẹ iwọn otutu nitorinaa eewu aiyipada 5% ti a sọ ni deede ni deede si 5% oṣuwọn aiyipada itan fun awọn awin idiyele.
Nẹtiwọọki iwadii aisan iṣoogun jẹ atunṣe pẹlu isotonic ifasẹyin nitoribẹẹ “iṣeeṣe giga ti arun” ṣe afihan isẹlẹ otitọ ṣaaju ki awọn oniwosan ṣiṣẹ.
Akopọ iwo wiwakọ ti ara ẹni ṣe iwọn igbẹkẹle-iwari ohun nitoribẹẹ 90% Dimegilio arinkiri ni igbẹkẹle ni deede nipasẹ module igbero.
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 Probability Calibration 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
Igbẹkẹle AI Trust
Awọn ibeere ti a beere nigbagbogbo
What is Probability Calibration?
Isọdiwọn tumọ si awọn iṣeeṣe ti a sọ ti awoṣe baramu otitọ: nigbati o sọ 70%, iṣẹlẹ yẹ ki o ṣẹlẹ nipa 70% ti akoko naa. O ṣe pataki nitori igbẹkẹle deede n ṣe awọn ipinnu to dara ni oogun, iṣuna, ati AI ti o ni eewu.
Kini iwọn iwọn otutu ṣe si awọn abajade ti nẹtiwọọki nkankikan?
Iwọn iwọn otutu pin awọn logits nipasẹ T kan ti o kọ ẹkọ ati tun-ṣiṣẹ softmax, rirọ tabi awọn iṣeeṣe didasilẹ laisi iyipada argmax.
Kini apẹrẹ ti o ni igbẹkẹle ṣe idite?
Aworan ti o gbẹkẹle ṣe afiwe igbẹkẹle asọtẹlẹ si igbohunsafẹfẹ abajade gangan; akọ-rọsẹ duro fun isọdiwọn pipe.
Kini idi ti awoṣe ti o peye ga julọ tun le nilo isọdiwọn?
Awoṣe le ṣe ipo awọn asọtẹlẹ daradara (ipeye giga) sibẹsibẹ sọtọ awọn iye iṣeeṣe ti ko baramu, nitorinaa iwọnwọn gbọdọ jẹ ṣayẹwo lọtọ.