I-AI Yezinhlelo Ezifakiwe: Kungani Ishintsha Konke

I-AI yayihlala kumaseva amakhulu nama-GPU amafu. Manje iyancipha futhi ishelela eduze kwezinzwa. I-AI yezinhlelo ezifakiwe ayithembisi kangako - isivele izwakala ngaphakathi kwamafriji, ama-drone, izinto ezigqokwayo ... ngisho namadivayisi angabukeki “ehlakaniphile” nhlobo.

Nasi isizathu esenza lolu shintsho lube lubalulekile, yini eyenza kube nzima, nokuthi yiziphi izinketho ezifanele isikhathi sakho.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Amathuluzi okuphatha i-AI amahle kakhulu aqinisekisa izinhlelo ze-AI ezithobela imithetho yokuziphatha nezisobala
Umhlahlandlela wamathuluzi asiza ukugcina i-AI enokuziphatha okuhle, ethobela imithetho, nesobala.

🔗 Isitoreji sezinto ze-AI: izinketho, izinketho, izinketho
Ukuqhathaniswa kwezinketho zokugcina izinto ezenzelwe imithwalo yemisebenzi ye-AI.

🔗 Izidingo zokugcina idatha ze-AI: lokho okudingeka ukwazi ngempela
Izinto ezibalulekile okufanele uzicabangele lapho uhlela ukugcinwa kwedatha ye-AI.

I-AI Yezinhlelo Ezifakiwe🌱

Amadivayisi afakiwe mancane, avame ukusebenza ngebhethri, futhi anezisetshenziswa eziningi. Kodwa i-AI ivula izinzuzo ezinkulu:

• Izinqumo zesikhathi sangempela ngaphandle kohambo oluya nokubuya lwamafu.

• Ubumfihlo ngomklamo - idatha eluhlaza ingahlala kudivayisi.

• Yehlisa ukubambezeleka lapho ama-millisecond ebalulekile.

• Ukuqagela okuqaphela amandla ngokusebenzisa imodeli ecophelelayo + izinketho zehadiwe.

Lezi akuzona izinzuzo ezihambisana nesandla: ukusunduza i-compute emaphethelweni kunciphisa ukuncika kwenethiwekhi futhi kuqinisa ubumfihlo ezimweni eziningi zokusetshenziswa [1].

Icebo alikho amandla amakhulu—kuwukuhlakanipha ngezinsiza ezilinganiselwe. Cabanga ngokugijima umjaho we-marathon unesikhwama semali… onjiniyela baqhubeka nokususa izitini.

Ithebula Lokuqhathanisa Okusheshayo le-AI Yezinhlelo Ezifakiwe 📝

(Yebo, ukungalingani. Kunjalo nangeqiniso.)

Kungani i-AI kumadivayisi afakiwe ibalulekile embonini 🏭

Akuyona nje i-hype: emigqeni yefektri, amamodeli amancane athola amaphutha; kwezolimo, ama-node anamandla aphansi ahlaziya inhlabathi ensimini; ezimotweni, izici zokuphepha azikwazi "ukushaya ucingo ekhaya" ngaphambi kokubhuleka. Lapho ukubambezeleka kanye nobumfihlo kungaxoxiswana ngakho, ukuhambisa i-compute emaphethelweni kuyindlela ehlakaniphile [1].

I-TinyML: Iqhawe Elithule le-AI Ehlanganisiwe 🐜

I-TinyML isebenzisa amamodeli kuma-microcontroller anama-kilobytes kuya kuma-megabytes ambalwa e-RAM - kodwa isakwazi ukubona amagama angukhiye, ukuqashelwa kwezimpawu, ukutholwa kwezimpawu, nokuningi. Kufana nokubuka igundane liphakamisa isitini. Kuyanelisa ngendlela emangalisayo.

Imodeli yengqondo esheshayo:

• Izinyathelo zedatha: okokufaka okuncane, kwezinzwa zokusakaza.

• Amamodeli: ama-CNN/ama-RNN amancane, i-ML yakudala, noma amanethi ahlanganisiwe/alinganisiwe.

• Izabelomali: ama-milliwatts, hhayi ama-watts; KB–MB, hhayi i-GB.

Izinketho zehadiwe: Izindleko vs. Ukusebenza ⚔️

Ukukhetha ihadiwe yilapho amaphrojekthi amaningi eshintshashintsha khona:

• Ikilasi le-Raspberry Pi: i-CPU enobungane, esetshenziswa kabanzi; iqinile kuma-prototypes.

• I-NVIDIA Jetson: amamojula e-AI akhiwe ngenhloso (isib. i-Orin) aletha amashumi kumakhulu ama-TOPS ukuze kube nombono oqinile noma ama-stack amamodeli amaningi - kuhle, kodwa kuyabiza futhi kunamandla kakhulu [4].

• I-Google Coral (Edge TPU): i-ASIC accelerator eletha ama-TOPS angu-4 cishe ku-2W (~2 TOPS/W) kumamodeli alinganisiwe - ukusebenza kahle kakhulu uma imodeli yakho ifanelana nemingcele [3].

• Ama-Smartphone SoC (Snapdragon): athunyelwa ngama-NPU nama-SDK ukuze asebenzise amamodeli kahle kudivayisi.

Umthetho oyinhloko: ibhalansi yezindleko, ama-thermal, kanye nokubala. "Kuhle ngokwanele, yonke indawo" kuvame ukudlula "osezingeni eliphezulu, akukho ndawo."

Izinselele Ezivamile ku-AI Yezinhlelo Ezifakiwe 🤯

Onjiniyela bavame ukulwa nalokhu:

• Inkumbulo eqinile: amadivayisi amancane awakwazi ukusingatha amamodeli amakhulu.

• Isabelomali sebhethri: yonke i-milliamp ibalulekile.

• Ukulungiswa kwemodeli:

  • Ukulinganisa → isisindo/ukusebenza okuncane, okusheshayo kwe-int8/float16.

  • Ukuthena → susa izisindo ezingabalulekile ukuze kutholakale ukuncipha.

  • Ukuhlanganisa/ukwabelana ngesisindo → cindezela kakhulu.
    Lezi izindlela ezijwayelekile zokusebenza kahle kudivayisi [2].

• Ukwandisa: iklasi yokubonisa i-Arduino ≠ uhlelo lokukhiqiza izimoto olunemingcele yokuphepha, ukuphepha, kanye nomjikelezo wokuphila.

Ulungisa amaphutha? Cabanga ufunda incwadi ngembobo yokhiye… ugqoke amagilavu.

Izicelo Eziwusizo Ozozibona Okuningi Maduze 🚀

• Izinto ezigqokwayo ezihlakaniphile ezenza ulwazi lwezempilo kudivayisi.

• Amakhamera e-IoT ahlaba umkhosi ngemicimbi ngaphandle kokusakaza izithombe ezingavuthiwe.

• Abasizi bezwi abangaxhunyiwe ku-inthanethi bokulawula okungenazandla - akukho ukuthembela efwini.

• Ama-drone azimele okuhlola, ukulethwa, kanye nokunemba kwe-ag.

Ngamafuphi: I-AI isondela ngokoqobo - ezihlakaleni zethu, emakhishini ethu, nakuzo zonke ingqalasizinda yethu.

Indlela Abathuthukisi Abangaqala Ngayo 🛠️

1. Qala nge -TensorFlow Lite yokusetshenziswa kwamathuluzi abanzi kanye ne-MCU→ ukumbozwa kweselula; sebenzisa i-quantization/pruning kusenesikhathi [2].

2. Hlola i-ExecuTorch uma uhlala ezweni le-PyTorch futhi udinga isikhathi sokusebenza esilula kudivayisi kuselula naku-embedded [5].

3. Zama ama-Arduino + TinyML kits ukuze uthole ama-prototyping asheshayo futhi ajabulisayo.

4. Uthanda amapayipi abonakalayo? I-Edge Impulse yehlisa isithiyo ngokuthwebula idatha, ukuqeqesha, kanye nokusabalalisa.

5. Phatha ihadiwe njengesakhamuzi sesigaba sokuqala - isibonelo kuma-CPU, bese uqinisekisa ku-accelerator yakho eqondiwe (i-Edge TPU, i-Jetson, i-NPU) ukuqinisekisa ukubambezeleka, ama-thermals, kanye nama-deltas okunemba.

I-Mini-vignette: Ithimba lithumela i-vibration-anomaly detector ku-coin-cell sensor. Imodeli ye-float32 ayisebenzisi isabelomali samandla; i-int8 quantization inciphisa amandla ngokwesilinganiso, inciphisa inkumbulo, futhi i-duty-cycling i-MCU iqeda umsebenzi - akukho nethiwekhi edingekayo [2,3].

Uguquko Oluthule lwe-AI Yezinhlelo Ezifakiwe 🌍

Amaprosesa amancane, angabizi kakhulu afunda ukuzwa → ukucabanga → ukwenza - endaweni. Impilo yebhethri izohlala isikhathaza, kodwa indlela icacile: amamodeli aqinile, ama-compiler angcono, ama-accelerator ahlakaniphile. Umphumela? Ubuchwepheshe obuzwakala buyimfihlo futhi buphendula ngoba abuxhunyiwe nje kuphela - bunaka.

Izinkomba

[1] I-ETSI (i-Multi-access Edge Computing) - Izinzuzo zokubambezeleka/zobumfihlo kanye nomongo wemboni.
I-ETSI MEC: Ukubuka konke kwePhepha Elimhlophe Elisha

[2] I-Google TensorFlow Model Optimization Toolkit - Ukulinganisa, ukuthena, ukuhlanganisa ukuze kusebenze kahle kudivayisi.
Umhlahlandlela Wokwenza Kahle Imodeli ye-TensorFlow

[3] I-Google Coral Edge TPU - Amabhentshimakhi e-Perf/W okusheshisa umkhawulo.
Amabhentshimakhi e-Edge TPU

[4] I-NVIDIA Jetson Orin (Esemthethweni) - Amamojula e-Edge AI kanye nezimvilophu zokusebenza.
Ukubuka Konke Kwamamojula e-Jetson Orin

[5] I-PyTorch ExecuTorch (Amadokhumenti Asemthethweni) - Isikhathi sokusebenza se-PyTorch kudivayisi yeselula kanye ne-edge.
Ukubuka Konke kwe-ExecuTorch

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI