Edge AI ni iki?

Edge AI ishyira ubwenge ahantu amakuru avuka. Bisa nkaho ari byiza, ariko igitekerezo cy'ingenzi ni cyoroshye: tekereza neza iruhande rwa sensor kugira ngo ibisubizo bigaragare ubu, ntabwo ari nyuma. Ubona umuvuduko, ukwizerwa, n'inkuru nziza y'ibanga nta gufata icyemezo cyose. Reka tubisobanure - harimo inzira ngufi n'ibindi bibazo byo ku ruhande. 😅

Ingingo ushobora gukunda gusoma nyuma y'iyi:

🔗 Igikoresho cyo gukora imibonano mpuzabitsina gitanga umusaruro (AI) ni iki?
Ibisobanuro bisobanutse neza ku buryo bwo gukora imibonano mpuzabitsina ikora neza, uko ikora, n'uburyo ikoreshwa mu buryo bufatika.

🔗 Ubuhanga bwo gukora imibonano mpuzabitsina (AI) ni iki?
Incamake y'ubuhanga bwo gukora ibintu mu buryo bwa "AI", imyitwarire y'ubwikorezi, n'imiterere y'ikoreshwa ry'ibintu mu buryo bufatika.

🔗 Uburyo bwo gupima ubushobozi bwa AI ni iki?
Menya uburyo bwo kwagura sisitemu za AI mu buryo bwizewe, bunoze, kandi buhendutse.

🔗 Porogaramu ya mudasobwa ya AI ni iki?
Isesengura ry'imiterere ya porogaramu za AI, inyungu z'ubwubatsi, n'iby'ibanze mu ishyirwa mu bikorwa.

Edge AI ni iki? Igisobanuro cyihuse 🧭

Edge AI ni uburyo bwo gukoresha moderi zatojwe zo kwigira imashini ku bikoresho bikusanya amakuru, kamera, robots, imodoka, ibikoresho byo kwambara, ibikoresho byo kugenzura inganda, nk'uko wabitekereza. Aho kohereza amakuru asanzwe kuri seriveri za kure kugira ngo zisesengurwe, igikoresho gitunganya amakuru mu gace runaka kandi cyohereza incamake gusa cyangwa ntayo kigize. Ingendo nke zo kugaruka no kugaruka, gutinda guke, kugenzura gukomeye. Niba ushaka ibisobanuro bisukuye kandi bidashingiye ku mucuruzi, tangira hano. [1]

Ni iki gituma Edge AI iba ingirakamaro mu by'ukuri? 🌟

• Gutinda guke - ibyemezo bibaho kuri mudasobwa, bityo ibisubizo bisa n'aho ari byo byihuse ku mirimo yo kubona ibintu nko kumenya ibintu, kumenya amagambo atunguranye, cyangwa gutanga integuza zitari nziza. [1]

• Ubuzima bwite bushingiye ku gace runaka - amakuru y'ingenzi ashobora kuguma kuri terefone, akagabanya uburyo agaragara kandi agafasha mu biganiro byo kugabanya amakuru. [1]

• Kubika bandwidth - kohereza ibintu cyangwa ibikorwa aho kohereza inkuru mbisi. [1]

• Ubushobozi bwo kwihangana - bikora neza mu gihe cyo guhuza ibintu mu buryo budasobanutse neza.

• Kugenzura ikiguzi - imikorere mito yo kubara mu bicu no kugabanuka kw'isohoka ry'amakuru.

• Kumenya aho ibintu biherereye - igikoresho "kiyumva" ibidukikije kandi kikagenda neza.

Inkuru yihuse: umupilote w’ubucuruzi yahinduye kamera zihoraho zoherezwa mu byiciro by’umuntu ugereranije n’ibintu biri kuri mashini, maze yongera kubara amasaha n’amashusho y’ibidasanzwe gusa. Ibisubizo: imenyesha rya ms munsi ya 200 ku nkengero z’ububiko n’igabanuka rya ~ 90% ry’urujya n’uruza rw’abasura urubuga - nta guhindura amasezerano ya WAN y’iduka. (Uburyo: igitekerezo cyo mu gace, gukusanya ibikorwa, ibintu bidasobanutse gusa.)

Edge AI vs cloud AI - itandukaniro ryihuse 🥊

• Aho kubara bibera : edge = on-device/hafi y'igikoresho; cloud = remote data centers.

• Ubukererwe : impera ≈ igihe nyacyo; igicu gifite ingendo zo kuzenguruka.

• Ihindagurika ry'amakuru : filters/compresses mbere na mbere; cloud ikunda kohereza amakuru yose uko yakabaye.

• Kwizerwa : edge ikomeza gukora nta interineti; cloud ikeneye guhuza.

• Imiyoborere : edge ishyigikira kugabanya amakuru; cloud ishyira mu bikorwa kugenzura. [1]

Si kimwe cyangwa bibiri. Sisitemu z'ubwenge zihuza byombi: ibyemezo byihuse byo mu gace utuyemo, isesengura ryimbitse n'inyigisho z'ubwikorezi mu buryo bw'ikoranabuhanga. Igisubizo cy'uruhererekane kirarambiranye kandi ni cyo.

Uburyo Edge AI ikora mu by'ukuri munsi y'ijisho 🧩

1. Utumashini dufata ibimenyetso by’umwimerere - amajwi, pixel za kamera, taps za IMU, n’ibimenyetso by’ihindagurika ry’amajwi.

2. Gutegura mbere y'uko ibintu bihinduka imiterere y'ibyo bimenyetso mu buryo bworoshye.

3. Igihe cyo gukoresha icyerekezo gikoresha icyitegererezo gito kuri igikoresho hakoreshejwe accelerators iyo kiboneka.

4. Gutunganya nyuma y'ibikorwa bihindura ibyavuye mu bikorwa mu bikorwa, ibirango, cyangwa ibikorwa byo kugenzura.

5. Telemetry ishyiramo gusa ibintu by'ingirakamaro: incamake, ibintu bidasanzwe, cyangwa ibitekerezo by'igihe runaka.

Amasaha yo gukoresha kuri mudasobwa uzabona mu ishyamba arimo Google's LiteRT (yahoze ari TensorFlow Lite), ONNX Runtime , na Intel's OpenVINO . Izi toolchains zigabanya umusaruro uva mu ngengo y'imari iciriritse/ububiko bw'ububiko hamwe n'amayeri nka quantization na operator fusion. Niba ukunda uburyo bwo gukoresha mudasobwa, inyandiko zazo zirakomeye. [3][4]

Aho bigaragara - aho ushobora gukoresha nyabyo 🧯🚗🏭

• Iyerekwa ku nkengero : kamera zo ku muryango (abantu ugereranije n'amatungo), gusuzuma ibikoresho byo mu nzu mu maduka, indege zitagira abapilote zigaragaza inenge.

• Amajwi kuri terefone : amagambo yo gukangura, amagambo yo kwandikira, kumenya amazi asohoka mu bimera.

• IoT y'inganda : moteri na pompe bigenzurwa kugira ngo harebwe niba hari ibibazo byo guhindagura mbere yuko bigwa.

• Imodoka : kugenzura abashoferi, kumenya inzira, gufasha guparika imodoka - munsi y'amasegonda cyangwa ku kibuno.

• Ubuvuzi : ibikoresho byo kwambara bigaragaza arrhythmias mu gace utuyemo; incamake zo guhuza nyuma.

• Telefoni zigendanwa : kunoza amafoto, kumenya abahamagara ubutumwa butari ubwa interineti, ibihe byo "telefone yanjye yabikoze ite utari kuri interineti".

Kubisobanuro byemewe (n'ikiganiro cy'umuvandimwe "igihu vs uruziga"), reba icyitegererezo cya NIST. [2]

Ibikoresho bituma irushaho kuba nziza 🔌

Hari urubuga ruke rugenzurwa cyane n'amazina yarwo:

• NVIDIA Jetson - Moduli zikoresha GPU za robots/cameras-Swiss-Army-knife vibes kuri AI iri muri mudasobwa.

• Google Edge TPU + LiteRT - uburyo bwo gusesengura umubare w'intege nke bunoze kandi bunoroshye bwo gukora imishinga ifite ingufu nke cyane. [3]

• Moteri y'imitsi ya Apple (ANE) - ML ikoreshwa cyane kuri iPhone, iPad, na Mac; Apple yasohoye akazi gafatika ko gushyira transformers neza kuri ANE. [5]

• CPU za Intel/iGPU/NPU hamwe na OpenVINO - "andika rimwe, shyira ahantu hose" kuri mudasobwa za Intel; uburyo bwo kunoza porogaramu buranyura.

• ONNX Runtime ahantu hose - igihe cyo gukora kidahinduka gifite abatanga serivisi zo gukoresha porogaramu zikoreshwa kuri telefoni, mudasobwa, n'amarembo. [4]

Ese byose ubikeneye? Oya rwose. Hitamo inzira imwe ikomeye ijyanye n'amato yawe maze ukomeze - churn ni umwanzi w'amakipe yashyizwemo.

Porogaramu ishyirwa ku rutonde - urugendo rugufi 🧰

• Uburyo bwo gukandagira : gupima (akenshi kugeza kuri int8), gukata, gusya.

• Kwihutisha urwego rw'umukoresha : imbuto zihujwe na silicon yawe.

• Igihe cyogukora : LiteRT, ONNX Runtime, GufunguraVINO. [3] [4]

• Ibipfunyika byo gushyiraho porogaramu : kontineri/amapaki ya porogaramu; rimwe na rimwe serivisi nto ku marembo.

• MLOps zo ku ruhande : Kuvugurura moderi ya OTA, gutanga A/B, imiyoboro ya telemetry.

• Ibikoresho bigenzura ubuzima bwite n'umutekano : uburyo bwo gufunga amakuru kuri mudasobwa, uburyo bwo kuyifungura neza, uburyo bwo kwemeza amakuru, n'ibindi bikoresho.

Isanduku nto: itsinda ry’abagenzuzi ryakoresheje drone ryakoresheje icyuma gipima ibintu bikomeye mu buryo bwa elegitoroniki bw’abanyeshuri basanzwe bakoresha LiteRT, hanyuma rihuza NMS kuri terefone. Igihe cyo gukora ingendo cyazamutse ~15% bitewe no kugabanuka kw’amakuru akoreshwa mu kubara; gushyira ingano yagabanutse ku mashusho adasanzwe. (Uburyo: gufata amakuru aho aherereye, gupima nyuma y’ingano, A/B yo mu gicucu mbere yo kuyashyira ahagaragara yose.)

Imbonerahamwe yo kugereranya - amahitamo azwi cyane ya Edge AI 🧪

Ibiganiro nyabyo: iyi meza ishingiye ku bitekerezo kandi irimo akajagari gato - kimwe n'isi nyayo.

Uburyo bwo guhitamo inzira ya Edge AI - igiti gito cyo gufata ibyemezo 🌳

• Ubuzima bwawe bugoranye mu gihe nyacyo? Tangira ukoresheje accelerators + models za quantized.

• Ubwoko bwinshi bw'ibikoresho? Hitamo ONNX Runtime cyangwa OpenVINO kugira ngo bikoreshwe neza. [4]

• Kohereza porogaramu igendanwa? LiteRT ni inzira igoye cyane. [3]

• Robotike cyangwa isesengura rya kamera? Uburyo bwa Jetson bwo gukoresha GPU buzigama umwanya.

• Uburyo bwo kubika amakuru mu ibanga buhamye? Komeza ushyire amakuru mu mwanya wawe, uyashyire mu ibanga igihe cyose ukiri mu mwanya usanzwe, aho kuyashyira mu mwanya usanzwe.

• Ikipe nto? Irinde imiyoboro y'ibikoresho idasanzwe - birambirana ni byiza.

• Imiterere izahinduka kenshi? Teganya OTA na telemetry guhera ku munsi wa mbere.

Ingaruka, imipaka, n'ibice birambirana ariko by'ingenzi 🧯

• Uburyo bwo guhindagurika - ibidukikije birahinduka; kurikirana uko ibintu bikwirakwizwa, koresha uburyo bwo gushushanya, kandi wongere witoze buri gihe.

• Gupima uburebure bw'ibikoresho - ububiko buke/imbaraga z'amashanyarazi mato cyangwa uburyo bworoshye bworoshye.

• Umutekano - fata uburyo bwo kwinjira; koresha boot ifite umutekano, ibikoresho byashyizweho umukono, icyemezo, serivisi zidafite amahirwe menshi.

• Imicungire y'amakuru - gutunganya amakuru mu gace utuyemo birafasha, ariko uracyakeneye uburenganzira, kubika amakuru, no gukoresha uburyo bwa "scoped telemetry".

• Fleet ops - ibikoresho bihagarara igihe bitari kuri interineti mu bihe bibi cyane; imiterere y'amavugurura yasubitswe n'ibyoherezwa bishobora kongera gushyirwa kuri interineti.

• Itsinda ry'abahanga mu ikoranabuhanga - ryashyizwemo + ML + DevOps ni itsinda ry'abantu batandukanye; bakora imyitozo ya buri munsi hakiri kare.

Gahunda ifatika yo kohereza ikintu cy'ingirakamaro 🗺️

1. Hitamo agasanduku kamwe gakoreshwa gafite agaciro gapima ku murongo wa 3, ijambo rya "wake word" kuri "smart speaker", nibindi.

2. Kusanya amakuru atunganye agaragaza aho ibintu biherereye; shyiramo urusaku kugira ngo ruhuze n'ukuri.

3. Ishusho iri ku gikoresho cy’ikoranabuhanga kiri hafi y’ibikoresho byo gukora.

4. Kanda icyitegererezo ukoresheje uburyo bwo gupima/gukata; pima igihombo cy'ukuri mu buryo butaziguye. [3]

5. Shyiraho igitekerezo muri API isukuye hamwe n'umuvuduko w'umugongo n'abashinzwe kurinda - kuko ibikoresho bimanitse saa mbiri za mu gitondo

6. Kora telemetri yubaha ubuzima bwite: imibare yo kohereza, histogramu, ibintu byakuwe mu mpande.

7. Komeza umutekano : binari zasinyweho, boot irinda umutekano, serivisi nke zifunguye.

8. Gahunda ya OTA : gukwirakwiza ibintu mu buryo butajegajega, gukwirakwiza ibintu mu buryo bw'izuba, gusubiza ibintu mu buryo bwihuse.

9. Umupilote ari mu gasanduku k'imfuruka gakomeye mbere na mbere - niba arokotse aho ngaho, azarokoka aho ari ho hose.

10. Gupima ukoresheje igitabo cy'imikino : uburyo wongeramo moderi, guhinduza imfunguzo, kubika amakuru - bityo umushinga #2 ntabwo ari akajagari.

Ibibazo Bikunze Kubazwa - ibisubizo bigufi kuri Curiosities ni iki

Ese Edge AI ikoresha moderi nto kuri mudasobwa nto?
Ahanini, yego - ariko ingano si yo nkuru yose. Nanone ireba ingengo y'imari yo gutinda, amasezerano y'ubuzima bwite, no gutegura ibikoresho byinshi bikorera mu gace runaka ariko biga ku isi hose. [1]

Ese nshobora no kwitoza ku rundi ruhande?
Hari imyitozo yoroheje/igenamiterere ryihariye ku bikoresho; imyitozo ikomeye iracyakorwa hagati. ONNX Runtime itanga amahitamo yo kwitoza ku bikoresho niba uri umuntu ukunda kwinezeza. [4]

Edge AI vs. Gukoresha imashini mu gupima ibihu ni iki?
Igihu na edge ni ibihuzwa. Byombi byongera ikoranabuhanga ryo kubara hafi y'amasoko y'amakuru, rimwe na rimwe binyuze mu nzira ziri hafi aho. Kugira ngo ubone ibisobanuro byemewe n'imiterere y'ikoranabuhanga, reba NIST. [2]

Ese Edge AI ihora inoza ubuzima bwite?
Birafasha - ariko si amayobera. Uracyakeneye kugabanya, inzira zihamye zo kuvugurura, no kwandika witonze. Fata ubuzima bwite nk'akamenyero, ntabwo ari agasanduku ko kwibandaho.

Gusoma cyane ushobora kuba warasomye 📚

1) Gutunganya moderi bidasenya uburyo ikora neza

Gupima ibintu bishobora kugabanya kwibuka no kwihutisha imikorere, ariko gupima hakoreshejwe amakuru agaragara cyangwa icyitegererezo bishobora gutuma squirrels zireba ahantu hari traffic cones. Gutunganya ibintu mu buryo bwa distillation - umwarimu uyobora umunyeshuri muto - akenshi bibungabunga imiterere y'ibintu. [3]

2) Igihe cyo gukora isuzuma rya Edge mu myitozo

Umusemuzi wa LiteRT akora akazi ko gushakisha amakuru mu buryo butunguranye kandi nta gahinda gakomeye mu gihe cyo gukora. ONNX Runtime ikoresha accelerators zitandukanye binyuze mu batanga serivisi zo gukora. Nta n'umwe muri bo ukoresha imbunda ya feza; byombi ni inyundo zikomeye. [3][4]

3) Gukomera mu gasozi

Ubushyuhe, umukungugu, ingufu zidafite imbaraga, Wi-Fi ifunguye: kora ibikoresho byo kugenzura bitangira gukora imiyoboro, gufata ibyemezo byo kubika amakuru, no kwiyunga iyo umuyoboro ugarutse. Ariko ntibishimishije cyane nk'uko abantu babibona - ni ingenzi cyane.

Interuro uzasubiramo mu nama - Edge AI ni iki 🗣️

Edge AI yegereza amakuru kugira ngo ihuze n'imbogamizi zifatika zo gutinda, ubuzima bwite, bandwidth, no kwizerwa. Ubumaji si chip imwe cyangwa framework imwe - ni uguhitamo neza icyo ugomba kubara aho uri.

Amagambo ya nyuma - Ndamaze igihe kirekire cyane, sinayasomye 🧵

Edge AI ikoresha moderi hafi y'amakuru kugira ngo ibicuruzwa byumvikane vuba, byihariye, kandi bikomere. Uzavanga ibitekerezo byo mu gace utuyemo n'ibicu kugira ngo ubone ibyiza byombi. Hitamo igihe cyo gukora gihuye n'ibikoresho byawe, wishingikirize ku byuma bikurura igihe cyose bishoboka, komeza moderi zisukuye hamwe no gukanda, kandi ushushanye imikorere y'imodoka nk'uko akazi kawe gashingiyeho - kuko, birashoboka. Niba umuntu akubajije Edge AI ni iki , vuga: ibyemezo bisobanutse, bifatwa mu gace utuyemo, ku gihe. Hanyuma umwenyure hanyuma uhindure ikintu ukoresheje bateri. 🔋🙂

Amareferensi

1. IBM - Edge AI ni iki? (ibisobanuro, inyungu).
   https://www.ibm.com/think/topics/edge-ai

2. NIST - SP 500-325: Icyitegererezo cy'Ikoranabuhanga rya Fog Computing (imiterere yemewe y'igihu/inkombe).
   https://csrc.nist.gov/pubs/sp/500/325/final

3. Google AI Edge - LiteRT (yahoze ari TensorFlow Lite) (igihe cyo gukora, quantization, kwimuka).
   https://ai.google.dev/edge/litert

4. Igihe cyo gukora cya ONNX - Amahugurwa yo gukoresha kuri mudasobwa (igihe cyo gukora cyoroshye + imyitozo yo gukoresha kuri mudasobwa zo mu nkengero).
   https://onnxruntime.ai/docs/get-started/training-on-device.html

5. Ubushakashatsi ku Imashini za Apple - Gukoresha Transformers kuri Moteri y'Imitsi ya Apple (amakuru y'imikorere ya ANE).
   https://machinelearning.apple.com/research/neural-engine-transformers

Shaka ubuhanga bwa AI bugezweho mu iduka ryemewe rya AI Assistant Store

Ku bijyanye natwe