Machine Learning qauv muaj thoob plaws qhov chaw tam sim no. Thaum nruab hnub, tej zaum koj yuav siv cov qauv no ntau dua li koj paub. Cov qauv kev kawm tshuab tau siv rau hauv kev ua haujlwm xws li kev tshaj tawm hauv xov xwm, kev yees duab, thiab tshuaj xyuas huab cua.
Lub tshuab-kev kawm algorithm yuav tau pom zoo rau qhov blog no rau koj. Peb txhua tus tau hnov txog li cas nws siv sijhawm los cob qhia cov qauv no. Peb txhua tus tau hnov tias kev cob qhia cov qauv no yog siv sijhawm ntev.
Txawm li cas los xij, ua qhov kev xav ntawm cov qauv no feem ntau suav nrog nqi.
Peb xav tau cov tshuab computer uas ceev txaus los tswj tus nqi uas peb tab tom siv cov kev pabcuam kev kawm tshuab. Yog li ntawd, feem ntau ntawm cov qauv no yog khiav ntawm cov ntaub ntawv loj loj nrog CPU thiab GPU pawg (txawm tias TPUs hauv qee kis).
Thaum koj thaij duab, koj xav tau tshuab kev kawm txhawm rau txhim kho nws tam sim ntawd. Koj tsis xav kom yuav tsum tau tos kom cov duab xa mus rau qhov chaw khaws ntaub ntawv, ua tiav, thiab xa rov qab rau koj. Hauv qhov no, tus qauv kev kawm tshuab yuav tsum tau ua hauv zos.
Thaum koj hais "Hav Siri" lossis "OK, Google," koj xav kom koj cov khoom siv teb tam sim ntawd. Tos kom koj lub suab xa mus rau cov khoos phis tawj, qhov twg nws yuav raug soj ntsuam thiab cov ntaub ntawv tau txais.
Qhov no yuav siv sij hawm thiab muaj kev cuam tshuam rau cov neeg siv kev paub. Hauv qhov no, koj xav kom tus qauv kev kawm tshuab ua haujlwm hauv zos thiab. Qhov no yog qhov twg TinyML tuaj.
Hauv tsab xov xwm no, peb yuav saib rau hauv TinyML, nws ua haujlwm li cas, nws siv, yuav pib li cas, thiab ntau ntxiv.
Yuav ua li cas yog TinyML?
TinyML yog ib qho kev qhuab qhia zoo tshaj plaws uas siv lub peev xwm hloov pauv ntawm kev kawm tshuab rau qhov kev ua tau zoo thiab lub zog txwv ntawm cov khoom siv me me thiab cov kab ke.
Kev xa mus ua tiav hauv kev lag luam no yuav tsum muaj kev nkag siab zoo ntawm cov ntawv thov, algorithms, hardware, thiab software. Nws yog lub tshuab kev kawm subgenre uas siv cov kev kawm sib sib zog nqus thiab cov qauv kev kawm tshuab hauv cov tshuab embedded uas ntiav microcontrollers, digital signal processors, los yog lwm yam ultra-low-power tshwj xeeb processors.
TinyML-enabled embedded devices yog npaj los khiav lub tshuab kev kawm algorithm rau ib txoj haujlwm tshwj xeeb, feem ntau yog ib feem ntawm lub cuab yeej. ntug xam.
Yuav kom khiav tau ntau lub lis piam, hli, lossis xyoo tsis tau them rov qab lossis hloov lub roj teeb, cov tshuab embedded no yuav tsum muaj lub zog siv tsawg dua 1 mW.
Ua li cas nws ua hauj lwm?
Tsuas yog lub hauv paus kev kawm tshuab uas tuaj yeem siv nrog microcontrollers thiab computers yog TensorFlow Lite. Nws yog ib txheej ntawm cov cuab yeej uas cia cov neeg tsim khoom khiav lawv cov qauv ntawm lub xov tooj ntawm tes, kos, thiab cov khoom siv ntug, tso cai rau kev kawm tshuab ntawm ya.
Lub interface ntawm microcontroller yog siv los sau cov ntaub ntawv los ntawm cov sensors (xws li microphones, lub koob yees duab, lossis cov sensors embedded).
Ua ntej xa mus rau microcontroller, cov ntaub ntawv tau muab tso rau hauv huab-based tshuab kev kawm qauv. Batch kev cob qhia hauv hom offline feem ntau yog ua haujlwm los cob qhia cov qauv no. Cov ntaub ntawv sensor uas yuav siv rau kev kawm thiab kev xav twb tau txiav txim siab rau daim ntawv thov tshwj xeeb.
Yog tias tus qauv tau raug cob qhia kom pom cov lus tsa suab, piv txwv li, nws twb tau teeb tsa los tswj lub suab txuas ntxiv ntawm lub microphone.
Txhua yam twb tau ua tiav nrog kev pab ntawm huab platform zoo li Google Colab nyob rau hauv rooj plaub ntawm TensorFlow Lite, suav nrog kev xaiv dataset, normalization, underfitting lossis overfitting ntawm tus qauv, tsis tu ncua, cov ntaub ntawv augmentation, kev cob qhia, validation, thiab kev sim.
Tus qauv kawm tiav thaum kawg tau hloov pauv thiab xa mus rau microcontroller, microcomputer, lossis digital teeb liab processor tom qab kev kawm offline. Tus qauv tsis muaj kev cob qhia ntxiv tom qab tau tsiv mus rau ib qho khoom siv embedded. Hloov chaw, nws tsuas yog siv cov ntaub ntawv hauv lub sijhawm los ntawm cov sensors lossis cov khoom siv nkag los siv tus qauv.
Raws li qhov tshwm sim, TinyML tshuab kev kawm qauv yuav tsum muaj qhov tshwj xeeb ruaj khov thiab muaj peev xwm rov ua haujlwm tom qab xyoo lossis tsis tau rov ua dua. Tag nrho cov qauv muaj peev xwm underfitting thiab overfitting yuav tsum tau tshawb xyuas kom tus qauv tseem cuam tshuam rau lub sijhawm ntev, qhov zoo tshaj plaws tsis muaj qhov kawg.
Tab sis vim li cas thiaj siv TinyML?
TinyML tau pib ua ib qho kev mob siab rau tshem tawm lossis txo qis IoT qhov kev cia siab rau huab kev pabcuam rau kev ua haujlwm me me tshuab kev kawm kev ua haujlwm. Qhov no tsim nyog siv cov qauv kev kawm tshuab ntawm cov khoom siv ntawm lawv tus kheej. Nws muab cov txiaj ntsig tseem ceeb hauv qab no:
- Tsawg-zog noj: Daim ntawv thov TinyML yuav tsum zoo dua siv tsawg dua 1 milliWatt ntawm lub zog. Nrog xws li kev siv hluav taws xob qis, ib lub cuab yeej yuav txuas ntxiv muab cov lus xaus los ntawm cov ntaub ntawv sensor rau lub hlis lossis xyoo, txawm tias siv los ntawm lub roj teeb npib.
- Cov nqi qis dua: Nws yog tsim los khiav ntawm tus nqi qis 32-ntsis microcontrollers lossis DSPs. Cov microcontrollers no feem ntau yog ob peb xees txhua, thiab tag nrho cov embedded system tsim nrog lawv yog tsawg dua $ 50. Qhov no yog qhov kev xaiv zoo tshaj plaws rau kev ua haujlwm me me ntawm kev kawm tshuab ntawm qhov loj, thiab tshwj xeeb tshaj yog muaj txiaj ntsig zoo hauv IoT daim ntawv thov uas yuav tsum siv tshuab kev kawm.
- Tsawg latency: Nws daim ntawv thov muaj qis latency vim lawv tsis tas yuav thauj lossis pauv cov ntaub ntawv hauv lub network. Tag nrho cov ntaub ntawv sensor raug kaw hauv zos, thiab cov lus xaus yog kos siv tus qauv uas twb tau kawm lawm. Cov txiaj ntsig ntawm kev txiav txim siab yuav raug xa mus rau lub server lossis huab rau kev txiav lossis ua haujlwm ntxiv, txawm tias qhov no tsis yog qhov tseem ceeb rau lub cuab yeej ua haujlwm. Qhov no txo qis network latency thiab tshem tawm qhov xav tau ntawm kev kawm tshuab ua haujlwm ua haujlwm ntawm huab lossis server.
- Tsis pub twg paub: Nws yog ib qho kev txhawj xeeb loj hauv internet thiab nrog internet ntawm yam. Lub tshuab kev kawm ua haujlwm hauv TinyML apps tau ua hauv zos, tsis tas khaws lossis xa cov ntaub ntawv sensor / neeg siv mus rau server / huab. Yog li ntawd, txawm tias thaum txuas nrog lub network, cov ntawv thov no muaj kev nyab xeeb rau kev siv thiab tsis muaj kev pheej hmoo ntawm tus kheej.
daim ntawv sau npe
- Ua liaj ua teb - Thaum twg Cov neeg ua liaj ua teb thaij duab ntawm tsob ntoo, TensorFlow Lite daim ntawv thov kuaj pom muaj mob hauv nws. Nws ua haujlwm ntawm txhua lub cuab yeej thiab tsis xav tau kev sib txuas hauv internet. Cov txheej txheem tiv thaiv kev ua liaj ua teb nyiam thiab yog qhov tseem ceeb rau cov neeg ua liaj ua teb nyob deb nroog.
- Mechanics Maintenance - TinyML, thaum siv rau cov khoom siv qis, tuaj yeem txheeb xyuas qhov tsis zoo hauv lub tshuab. Nws suav nrog kev saib xyuas raws li kev kwv yees. Ping Services, Australian start-up, tau qhia txog IoT gadget uas saib xyuas cua turbines los ntawm kev txuas nws tus kheej mus rau lub turbine sab nraum. Nws ceeb toom rau tub ceev xwm thaum twg nws pom muaj teeb meem tshwm sim lossis ua haujlwm tsis zoo.
- Tsev Kho Mob – The Solar Scare yog ib qhov project. Cov yoov tshaj cum siv TinyML los tiv thaiv kev kis kab mob xws li dengue thiab malaria. Nws yog siv los ntawm lub hnub ci zog thiab tshawb pom cov yoov tshaj cum yug me nyuam ua ntej qhia cov dej kom txwv tsis pub yoov tshaj cum.
- Traffic Surveillance – By thov TinyML rau cov sensors uas sau cov ntaub ntawv tsheb khiav hauv lub sijhawm tiag tiag, peb tuaj yeem siv lawv los tswj cov tsheb khiav zoo dua thiab txiav lub sijhawm teb rau cov tsheb xwm txheej ceev. Swim.AI, piv txwv li, siv cov thev naus laus zis no ntawm cov ntaub ntawv streaming kom muaj kev nyab xeeb rau cov neeg caij tsheb thaum tseem txo qis kev sib tsoo thiab emissions los ntawm kev siv ntse.
- kev cai lij choj: TinyML tuaj yeem siv rau hauv tub ceev xwm txhawm rau txheeb xyuas qhov ua txhaum cai xws li kev tawm tsam thiab tub sab nyiag siv tshuab kev kawm thiab kev paub txog kev piav tes piav taw. Ib qho kev pab cuam zoo sib xws kuj tseem siv tau los tiv thaiv lub txhab nyiaj ATMs. Los ntawm kev saib tus neeg siv tus cwj pwm, tus qauv TinyML tuaj yeem kwv yees seb tus neeg siv puas yog tus neeg siv khoom tiag tiag ua tiav kev sib pauv lossis tus neeg tawm tsam sim nyiag lossis rhuav tshem lub ATM.
Yuav pib nrog TinyML li cas?
Txhawm rau pib nrog TinyML hauv TensorFlow Lite, koj yuav xav tau lub rooj tsav xwm microcontroller sib xws. TensorFlow Lite rau Microcontrollers txhawb cov microcontrollers teev hauv qab no.
- Wio Terminal: ATSAMD51
- Himax WE-I Plus EVB Endpoint AI Development Board
- STM32F746 Discovery cov khoom siv
- Adafruit EdgeBadge
- Synopsys DesignWare ARC EM Software Development Platform
- Sony Kev Tshaj Tawm
- Arduino Nano 33 BLE Sense
- SparkFun Ntug
- Adafruit TensorFlow Lite rau Microcontrollers Kit
- Adafruit Circuit Chaw Ua Si Bluefruit
- Espressif ESP32-DevKitC
- Espressif ESP-Qhov muag
Cov no yog 32-ntsis microcontrollers muaj txaus flash nco, RAM, thiab moos zaus los ua tus qauv kev kawm tshuab. Cov laug cam kuj muaj ib tug xov tooj ntawm onboard sensors muaj peev xwm khiav ib qho kev pab cuam embedded thiab siv tshuab kev kawm qauv rau lub hom phiaj daim ntawv thov. Rau tsim ib lub tshuab kawm qauv, koj yuav xav tau lub laptop lossis computer ntxiv rau lub hardware platform.
Txhua lub platform kho vajtse muaj nws tus kheej cov cuab yeej programming rau lub tsev, kev cob qhia, thiab porting tshuab kev kawm qauv, uas siv TensorFlow Lite rau Microcontrollers pob. TensorFlow Lite yog pub dawb rau siv thiab hloov kho vim nws yog Qhib qhov chaw.
Txhawm rau pib nrog TinyML thiab TensorFlow Lite, txhua yam koj xav tau yog ib qho ntawm cov saum toj no-hais embedded hardware platforms, lub computer / laptop, USB cable, USB-rau-Serial converter - thiab lub siab xav xyaum tshuab kev kawm nrog cov tshuab kos. .
Txoj kev sib tw
Txawm hais tias TinyML qhov kev nce qib tau ua rau muaj txiaj ntsig zoo, kev lag luam kev kawm tshuab tseem muaj kev cuam tshuam loj heev.
- Software muaj ntau haiv neeg – Tes-coding, cov cim cim, thiab cov neeg txhais lus ML yog txhua txoj kev xaiv rau kev xa cov qauv ntawm TinyML li, thiab txhua tus siv sijhawm sib txawv thiab siv zog. Kev ua yeeb yam sib txawv tuaj yeem tshwm sim los ntawm qhov no.
- Hardware diversity – Muaj muaj ntau yam kev xaiv kho vajtse muaj. TinyML platforms tuaj yeem yog txhua yam los ntawm cov hom phiaj dav dav microcontrollers mus rau cov txheej txheem neural. Qhov no ua rau muaj teeb meem nrog cov qauv xa tawm thoob plaws cov qauv sib txawv.
- Troubleshooting/debugging – Thaum twg tus qauv ML ua haujlwm tsis zoo ntawm huab, nws yooj yim los saib cov ntaub ntawv thiab txheeb xyuas qhov yuav ua li cas. Thaum tus qauv nthuav dav thoob plaws ntau txhiab ntawm TinyML cov khoom siv, tsis muaj cov ntaub ntawv rov qab mus rau huab, kev debugging yuav nyuaj thiab yuav xav tau ib txoj kev sib txawv.
- Memory constraints - Traditional platforms, xws li smartphones thiab laptops, xav tau gigabytes ntawm RAM, whereas TinyML li siv kilobytes los yog megabytes. Yog li ntawd, qhov luaj li cas ntawm tus qauv uas yuav siv tau yog txwv.
- Kev cob qhia qauv - Txawm tias muaj ntau qhov zoo rau kev xa cov qauv ML ntawm TinyML cov khoom siv, feem ntau ntawm ML qauv tseem raug cob qhia ntawm huab kom rov ua dua thiab txuas ntxiv txhim kho cov qauv raug.
Yav tom ntej
TinyML, nrog nws cov hneev taw me me, siv roj teeb tsawg, thiab tsis muaj lossis txwv kev cia siab rau kev sib txuas hauv internet, muaj peev xwm loj heev nyob rau yav tom ntej, raws li feem ntau ntawm nqaim. artificial txawj ntse yuav raug muab coj los siv rau ntawm cov khoom siv ntug lossis cov khoom siv ywj pheej embedded.
Nws yuav ua rau IoT daim ntawv thov ntiag tug thiab nyab xeeb los ntawm kev siv lawv. Txawm li cas los TensorFlow Lite tam sim no yog lub hauv paus kev kawm tshuab ib leeg rau microcontrollers thiab microcomputers, lwm yam piv txwv xws li sensor thiab ARM's CMSIS-NN yog nyob rau hauv cov hauj lwm.
Thaum TensorFlow Lite yog qhov qhib qhov project hauv kev ua tiav uas tau tawm mus rau qhov pib zoo heev nrog Google Pab Pawg, nws tseem xav tau kev txhawb nqa hauv zej zog kom nkag mus rau hauv lub ntsiab.
xaus
TinyML yog ib txoj hauv kev tshiab uas ua ke nrog cov tshuab embedded nrog kev kawm tshuab. Raws li qhov nqaim AI peaks hauv ntau qhov ntsug thiab cov thawj, cov thev naus laus zis tuaj yeem tshwm sim los ua qhov tseem ceeb hauv kev kawm tshuab thiab kev txawj ntse.
Nws muab kev daws teeb meem rau ntau yam kev cov nyom uas IoT sector thiab cov kws tshaj lij siv tshuab kev kawm rau ntau qhov kev qhuab qhia tshwj xeeb tam sim no tab tom ntsib.
Lub tswv yim ntawm kev siv tshuab kev kawm ntawm ntug cov khoom siv nrog me me suav hneev taw thiab kev siv hluav taws xob muaj peev xwm hloov pauv tau li cas cov kab ke thiab cov neeg hlau tau tsim.
Sau ntawv cia Ncua