Kwenzekeni
Ababhali bethula i-LiteEvent-AE, i-autoencoder engasindi, elungisekayo edizayinelwe ukuminyanisa idatha yokubonwayo esekwe kumcimbi ukuze kutholwe ukubambezeleka okuphansi kumadivayisi onqenqema akhawulelwe amandla. Babika ukunemba kokuncintisana noma okuphakeme ku-YOLOv9 kumadathasethi amabili ombono womcimbi, anamapharamitha ambalwa afika kwangu-35.6. Umthombo uphinde ubike ozimele abangu-44.8 ngesekhondi ku-NVIDIA Jetson Nano kanye nokusetshenziswa kwamandla okulinganiselwe okuphansi kakhulu ku-Raspberry Pi 4B ngaphansi kwephrothokholi yokuhlola yephepha.
Ikhasi le-arXiv langomhla zingama-22 ku-Aug. 2026 lichaza i-LiteEvent-AE njenge-autoencoder engasindi yombono osuselwe kumcimbi kumadivayisi aphethwe yi-low-latency, acindezelwe amandla. Inkinga eyinhloko yaleli phepha ukuthi ukusakazwa kwemicimbi akuvumelani futhi kunomsindo, kuyilapho amasistimu okufunda okujulile avamile angaba namandla kakhulu ngekhompyutha kumapulatifomu ashumekiwe anamandla aphansi. Umthombo wethula umsebenzi njengendlela ye-AI yokunciphisa lowo mthwalo ngenkathi igcina ulwazi oludingekayo ukuze uthole ukubonwa komfula.
Isistimu ehlongozwayo ihlanganisa isifaki khodi esingasindi esinomkhawulo womcimbi oguquguqukayo kanye nekhanda elincane lokuhlukanisa. Ngokusho kwe-abstract, i-autoencoder icindezela idatha ye-neuromorphic ngenkathi ilondoloza ukwakheka kwe-spatiotemporal ebalulekile. Umthombo awunikezi ngezakhiwo ezigcwele, inqubo yokuqeqesha, isibalo sepharamitha, inkumbulo yonyawo, noma imininingwane yokukhetha i-threshold, ngakho-ke i-abstract iyodwa ayikwazi ukucacisa ukuthi isistimu yenza kanjani ukuhwebelana kwayo noma ukuthi idizayini iphatheka kanjani kuyo yonke ihadiwe.
Ababhali babika ukuhlaziya kudathasethi yobuso be-Smart Event kanye ne-Event-Based Crossing Dataset. Ngokumelene ne-YOLOv9, bachaza i-LiteEvent-AE njengokuzuza ukunemba kokuncintisana noma okuphakeme kuyilapho isebenzisa amapharamitha afika kwangu-35.6 ambalwa. Awekho amanani anembile okunemba, ukuhlukaniswa kwezigaba, izikhawu zokuzithemba, noma ukuphindaphinda okuzimele okufakiwe kumthombo onikeziwe. Ngakho-ke lezi izimangalo ezenziwe ababhali be-preprint, hhayi okutholwe ngokuzimela ezintweni ezinikeziwe.
Leli phepha liphinde libike izivivinyo zehadiwe ku-Raspberry Pi 4B kanye ne-NVIDIA Jetson Nano. Ku-Jetson Nano, umthombo uthi i-LiteEvent-AE ifinyelele ozimele abangu-44.8 ngomzuzwana. Ku-Raspberry Pi 4B CPU, i-autoencoder engu-50% idle ama-joule angu-16.19 ngomthwalo ohloliwe we-inference, lapho ababhali ababala njengokusebenziseka kwamandla okuphansi cishe izikhathi ezingama-726.3 kune-YOLOv9 ngaphansi kwephrothokholi efanayo. Umthombo awucacisi ukuthi lesi silinganiso sihlanganisa inzwa, inkumbulo, isitoreji, noma ezinye izingxenye zesistimu.
Imininingwane yomthombo: arxiv.org ↗
Kungani kubalulekile
Amakhamera asekelwe kumcimbi akhiqiza amasiginali abonakalayo agqagqene, asynchronous, kodwa ukucubungula lezo zimpawu kungase kube nzima kumakhompyutha amancane, amandla aphansi. Uma imiphumela ebikiwe ibambe ngaphezu kokuhlolwa kwababhali, amamodeli ahlangene afana ne-LiteEvent-AE angenza ukuqashelwa okubonakalayo kwesikhathi sangempela kusebenze kakhulu ezinhlelweni ezishumekiwe nezamaselula lapho amamodeli avamile okubuka ngekhompyutha ebiza kakhulu. Ukuqhathaniswa kwamandla kungase kubaluleke kakhulu, kodwa kuwumphumela obikwe ngumlobi osuka emthwalweni othile womsebenzi futhi akufanele wenziwe okuvamile kukho konke ukuthunyelwa kwe-AI enqenqemeni.
Ukubaluleka okungokoqobo komsebenzi kuvela ekugxileni kwawo ekucabangeni onqenqemeni esikhundleni samasistimu ekhompuyutha amakhulu, amaphakathi nendawo. Umbono osuselwe kumcimbi uhloselwe ukunikeza amasiginali abonakalayo angencane, abambezeleke kancane, futhi imodeli engacubungula lezo zimpawu ku-hardware ehlangene inganciphisa isidingo sokudlulisa idatha noma ithembele kumakhompyutha anamandla kakhulu. Lokho kungabalulekile ezinhlelweni zeselula, ezizimele, nezishumekiwe ezihlonzwe ababhali, nakuba iphepha lingabonisi umkhiqizo osetshenzisiwe noma uhlelo lokusebenza.
Ukwehliswa kwepharamitha okubikiwe kuyahambisana ngoba usayizi wemodeli uthinta ngaphezu kwesitoreji. Amanethiwekhi amancane anganciphisa ukubala kanye nokucindezela kwenkumbulo, okuyizingqinamba ezivamile kumakhompyutha ebhodi elilodwa nezinye izinkundla ezishumekiwe. Umthombo uxhuma lokho kuphumelela ekusebenzeni kokuqashelwa, kodwa awutholi ukuthi i-LiteEvent-AE iqhathaniswa kanjani nesikhathi sokulayisha imodeli, ukusetshenziswa kwenkumbulo, ukuziphatha okushisayo, ukuphuma okuqinile, noma ukusebenza ngemva kokusebenza isikhathi eside.
Umphumela wamandla ungaba umphumela ikakhulukazi uma umngcele wayo wokulinganisa nephrothokholi yokuqhathanisa kumelela. Umehluko ophindwe izikhathi ezingu-726.3 ofunwayo ungashintsha ngokubonakalayo ukuba nokwenzeka kokusebenza kombono oqhubekayo ovela kubhajethi yamandla alinganiselwe. Nokho, isibalo siboshelwe kumthwalo owodwa ohloliwe kanye nokuqhathaniswa ne-YOLOv9. Kufanele ifundwe njengomphumela wokuhlola obikiwe, hhayi njengesilinganiso esijwayelekile semvelo noma izinzuzo zokusebenza ze-AI esekelwe kumcimbi.
Umsebenzi futhi ubonisa ukukhetha okubanzi kobunjiniyela: izethulo zedatha ezikhethekile zingavumela amasistimu amancane we-AI ukuthi ahlangabezane nezidingo zesikhathi sangempela ngaphandle kokumane kukale ihadiwe. Lokho kungenzeka kuwusizo kubacwaningi nakubathuthukisi abasebenza kumasistimu okubona abambekile. Noma kunjalo, okokusebenza okunikeziwe akubonisi ukuthi indlela ijwayeleka kwamanye amafomethi ekhamera yomcimbi, imisebenzi, indawo, noma izinqumo ezibalulekile zokuphepha, futhi ayiqinisekisi ukuthi amandla okucabanga aphansi akhiqiza ngokuzenzakalelayo amandla esistimu aphansi.
I-Interactive Mechanism: Indlela Esebenza Ngayo Ngempela
Hlola ubuchwepheshe obuyisisekelo ngemuva kwalokhu kuthuthukiswa ngokuhlanganyela.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Ongakubuka ngokulandelayo
Imibuzo eyinhloko ukuthi ingabe ukunemba okubikiwe, imiphumela, namandla akhiqizwa kabusha kuzo zonke izivivinyo ezizimele, amasethi wedatha, izinzwa zomcimbi, nezimo zokusebenza. Umthombo onikeziwe awunikezi izibalo zokunemba okunembile, osayizi bedathasethi, ukusatshalaliswa kwesikhathi eside, imingcele yokulinganisa amandla, noma ubufakazi bokusetshenziswa kwenkambu. Ukucutshungulwa okwengeziwe kufanele futhi kuhlole ukuthi umkhawulo womcimbi oguquguqukayo uziphatha kanjani ngaphansi komsindo, ukunyakaza okushintshwayo, nezimo ezihlukene zokukhanya, nokuthi ingabe ukusetshenziswa kuyatholakala ngendlela abanye abangaphinda bayikhiqize.
Okubalulekile kokuqala ukubuyekezwa okuseduze kwamathebula okuhlola aphelele ephepha. Abafundi kufanele babheke ukunemba kwangempela nezilinganiso zamaphutha kuwo womabili amasethi wedatha, ukulungiselelwa okunembile kwe-YOLOv9, ukukhetha kokucubungula kusengaphambili, izethulo zemicimbi, ukuhlukaniswa kwezigaba, kanye nokuthi wonke amasistimu athole ukushuna okufanayo. Umusho we-abstract othi "ukuncintisana noma uphakeme" akwanele ukuhlola usayizi noma ukwethembeka kwezibalo zomehluko wokusebenza okubikiwe.
Ukuphindaphinda okuzimele kungasiza ekunqumeni ukuthi ingabe imiphumela incike kumasethi edatha akhethiwe noma ihadiwe. Ukuhlola okuwusizo kungakala i-LiteEvent-AE emisebenzini eyengeziwe yokuqaphela esekelwe kumcimbi, izinzwa ezihlukene, namanye amaphrosesa anamandla aphansi kuyilapho kubikwa kokubili okumaphakathi nomsila. Abaphindaphindayo kufanele futhi babhale izisetshenziswa zamandla nemingcele yokulinganisa ukuze umphumela we-16.19-joule kanye nokuqhathanisa ne-YOLOv9 kuhunyushwe ngokungaguquki.
Ukuqina ngomunye umbuzo ovulekile. Umthombo ukhomba ukusakazwa kwemicimbi engavumelanisi nenomsindo njengezinselele futhi uthi indlela isebenzisa i-threshold eguquguqukayo, kodwa umbhalo onikeziwe awubiki ukuthi ukusebenza kushintsha kanjani ngomsindo wenzwa, i-threshold drift, ukunyakaza okusheshayo, imicimbi engatheni, noma ukushintsha ukukhanya. Lezo zimo zinganquma ukuthi ingabe imodeli ehlangene ihlala ithembekile ngaphandle kokuhlolwa kwedathasethi elawulwayo.
Okokugcina, abafundi kufanele babheke ubufakazi bokukhishwa okusebenzayo nokusetshenziswa: ikhodi ekhiqizwa kabusha noma amafayela emodeli, imiyalelo eqondene nehadiwe, ukuhlolwa kwemithwalo eqhubekayo, kanye nemibukiso eboshelwe emisebenzini yokuzimela yangempela noma yeselula. Umthombo usungula isiphakamiso socwaningo kanye nokuhlolwa okubikwe umbhali, kodwa awuqambi ukutholakala kwezentengiso, ukwethembeka kwenkambu, noma ukutholwa. Lokho kungaba intuthuko ehlukene kunemiphumela engacatshangelwa kusukela kulokhu kunyatheliswa kwangaphambili.