Lu xew
Auteur yi dañuy wane LiteEvent-AE, di autoencoder bu woyof, buñ mëna configurer, ñu defaree ko ngir kompresse ay done yu lalu ci xew-xew, ngir mëna am ay jafe-jafe yu néew ci aparey yu am energie bu bari. Dañuy wax ni YOLOv9 moo gëna dëggu wala gëna dëggu ci ñaari done yu ñuy gis ci xew-xew, ak lu ëpp 35.6 yoon lu gëna néew ay parametre. Woykat bi dafa wax itam ni 44.8 kadre ci segond bu nekk ci NVIDIA Jetson Nano ak lu gëna néew luñu natt ci Raspberry Pi 4B ci protokolu jàngat bu këyit bi.
Benn xëtu arXiv bu am bis 22 ut 2026 dafay leeral LiteEvent-AE ni autoencoder bu woyof ngir gis-gis bu lalu ci xew-xew ci aparey yu am latency bu woyof, yu am energie bu bari. Jafe-jafe bi gëna mag ci këyit bi mooy ni xew-xew yi dañu asynchrone ak bari bruit, ci noonu sistem yu xóot yi mën nañu nekk lu bari ci ordinatër ngir platform yu néew doole. Source bi dafay wane liggéey bi ni ab pexe IA ngir wàññi coono bi boole ci denc xibaar yiñ soxla ngir xàmmee downstream.
Sistem bi ñu nara def dafa boole encodeur convolutionnel bu woyof ak threshold buy méngoo ak boppu bu ndaw. Sunu sukkandikoo ci abstract bi, autoencoder bi dafay tënk done neuromorphique yi ci noonu lañuy baña yàq jumtukaayi spatiotemporal yu am solo yi. Source bi joxeewul architecture bi yépp, anam yi ñuy tàggatee, lim parametre yi, emprent mémoire bi, wala detay yiñ tànnee threshold, kon abstract bi kese mënul wane ni sistem bi di defee ay kompromis wala ni portable bi jëmmal ci hardware bi.
Auteur yi dañuy wax ci jàngat yi ñu def ci done yu Smart Event Face ak done yu lalu ci Xew-xew. Kontra YOLOv9, ñu ngi wax LiteEvent-AE ni defna njubte bu gëna mag wala bu gëna baax ci jëfandikoo lu ëpp 35.6 yoon lu néew ay parametre. Amul benn valeur bu dëggu, xaaj bu klaas, diggante wóolu, wala replication independent yuñ boole ci balluwaay biñ joxe. Kon loolu ay wax lañu yu bindkati preprint yi wax, duñu ay gis-gis yuñ joxe ci seen bopp ci mbir yiñ joxe.
Këyit dafay wax itam ni dañu natt ay aparey ci Raspberry Pi 4B ak NVIDIA Jetson Nano. Ci Jetson Nano, surnaal bi neena LiteEvent-AE yegg na 44.8 kaadar ci segond bu nekk. Ci CPU Raspberry Pi 4B, 50% autoencoder bi dafa lekk 16.19 joules ci liggéey biñ jàngat, bi bindkat yi xayma ni lu tollu ci 726.3 yoon la gëna néew energie ci YOLOv9 ci benn protocole bi. Boroom bi leeralul ndax natt gi dafa amaale kaptër bi, mémoire bi, dencukaay bi, wala yeneen mbir ci sistem bi.
Ay leeral ci cosaan: arxiv.org ↗
Lu tax mu am solo
Kamera yi lalu ci xew-xew yi dañuy defar siñaal yu bari te wuute, waaye liggéey siñaal yooyu mën na nekk lu jafe ci ordinatër yu ndaw te amul doole. Sudee liñu xamle ci resultaa yi weesuna test yi bindkat yi def, model yu kompact yu melni LiteEvent-AE mën nañu gëna xàmmee gis-gis ci jamono dëgg ci sistem yuñ samp ak mobile, fu model yu ordinatër-vision yi gëna yomb. Tegtale energie bi mën na am solo, waaye ab resultaa la bu bindkat bi xamle ci benn liggéey bu amul benn werante te waru ñu ko yamale ci bépp jëfandikoo edge-IA.
Njariñu liggéey bi ci jëfandikoo gi mingi aju ci liñuy bàyyi seen xel ci inference ci catu sistem ordinatër yu gëna mag, yuñ dajale. Li tax ñuy xool ci xew-xew mooy joxe siñaal yu bari te néew latency, ak model bu mëna jëfandikoo siñaal yooyu ci hardware bu kompact mën na wàññi soxla yónnee ay done wala nga yéem ci ordinatër yu gëna am doole. Loolu mën na am solo ci aplikaasioŋ mobile, autonome, ak yuñ samp yi bindkat yi xamme, ndigam këyit bi wanewul benn produit wala sistem operationel buñ dugal.
Wàññig parametre biñ xamle amna solo ndax dayo model bi moo gëna am njeexital ci dencukaay bi. Reseau yu ndaw yi mën nañu wàññi pressu ordinatër ak mémoire, te loolu mooy jafe-jafe bi gëna bari ci ordinatër yu am benn plaque ak ci yeneen platform yuñ samp. Source bi dafay boole efficacité bi ak performance reconnaissance, waaye mënul wane ni LiteEvent-AE di méngale ci diiru chargement model, jëfandikoo mémoire, jeffin thermique, produit bu yàgg, wala performance ginaaw liggéey bu yàgg.
Resultaa energie bi mën na nekk lu am solo sudee natt gi ak protokolu méngale bi dañuy wane. Différence bu 726.3-fold biñu wax ni dina soppi anam wi ñuy doxalee ci inference visuel bu wéy ci budget bu néew. Waaye, lim bi dafa lëkkaloo ak benn liggéey buñ jàngat ak méngale ak YOLOv9. Dañu ko wara jàppee ni njariñu jàngat buñ xamle, te bañ ko jàppee ni xayma bu ëpp solo ci njariñu environmaa bi wala liggéeyu IA bu sukkandiko ci xew-xew.
Liggéey bi dafay wane itam tanneef bu gëna yaatu ci wàllu xam-xam: joxe done yu yam mën na tax sistem IA yu ndaw yi mëna dëppoo ak bëgg-bëggu jamono dëgg te duñu yam ci yaatuwaayu aparey bi. Mën-mën boobu lu am solo la ci gëstukat yi ak defarkat yiy liggéey ci sistemu gis-gis bu tënk. Ba leegi, jumtukaay yiñ joxe wane wuñu ni anam wi dafay yamale yeneen formaa kamera yi, liggéey yi, environmaa yi, wala dogal yu am solo ci wàllu kaaraange, te wane wuñu ni energie inference bu gëna ndaw dafay defar ci saasi energie sistem bi gëna néew.
Mekanism buy weccoo xalaat: naka lay doxee
Saytu xarala yu bees yi ci ginaaw yokkute bii ci anam wu weccoo xalaat.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
Li nga wara seetaan ci topp
Laaj yi gëna am solo ñooy ndax njubte giñ siiwal, génne gi, ak njariñu energie yi dañuy bawoo ci test yu moom seen bopp, ensemble done, kaptëru xew-xew, ak anam yi ñuy doxalee. Boroom biñu joxe joxeewul ay lim yu leer, dayo done yi, séddaleb latency, ay pexe ngir natt doole, wala firnde ci jëfandikoo gi ci terrain bi. Yokkug saytu dafa wara xoolaat itam ni thresholding adaptive event di doxee ci biir bruit, coppite mouvement, ak anam yu wuute ci leeral, ak ndax amal gi am ci forme bu ñeneen ñi mëna joxe.
Li njëkk mooy xoolaat bu baax tablo experimentaal yi ci këyit bi. Jàngatkat yi dañu wara seet njub ak njuumte ci ñaari done yi, configuration YOLOv9 bu dëggu, tànneef yi ñuy njëkka def, misaali xew-xew yi, xaaj , ak ndax sistem yépp jot nañu tuning bu méngoo. Baatu abstract bi "compétitif wala superieur" doyul ngir jàngat dayo wala wóortéef ci lim yi ci wuute giñ xamle ci liggéey bi.
Replication moom boppam dina tax ñu xam ndax resultaa yi a ngi aju ci done yiñ tànn wala ci aparey bi. Test yu am njariñ dinañu natt LiteEvent-AE ci yeneen liggéey yu jëm ci xàmmee xew-xew, kaptër yu wuute, ak yeneen processeur yu néew doole, boole ci rapoor latency moyen ak geen. Replicateurs war nañu itam bind jumtukaayi doole ak natt ay pexe suko defee ñu mëna tekki 16.19-joule ak méngale ak YOLOv9 ci anam wu méngoo.
beneen laaj bu ubbeeku la. Source bi dafa xamme asynchrone ak bruit-prone xew-xew streams ni ay jafe-jafe te neena anam wi jëfandikoo thresholding adaptif, waaye bind biñ joxe waxul naka la performance di soppikoo ak bruit sensor, drift threshold, mouvement rapide, xew-xew yu néew, wala coppite leeral. Tegtal yooyu mën nañu wane ndax xeetu kompact bi mën na wéy di wóolu ci bitti jàngat yiñ saytu ci done yi.
Fi may jeexalee mooy, ñiy jàng war nañu xool ndax amna firnde ci génne ak jëfandikoo gi: kode wala fichier model yuñ mëna defaraat, tegtal yiñ jagleel hardware bi, test ci sargu liggéey yu yàgg, ak wane yu lëkkaloo ak liggéey yu autonome wala mobile. Boroom bi dafay tëral ab proposition de recherche ak ay experiment yu bindkat bi def, waaye du taxawal disponibilite ci njaay, wóor ci terrain bi, wala adoption. Loolu dina nekk ay jéego yu wuute, du ay njeexital yuñ mëna xalaat ci preprint bii.