Chii chaitika
Chinyorwa chearXiv chakatumirwa muna Aug. 28 chinosuma SpikeOPD, nzira yekugadzirisa vadzidzisi veneural network vakadzidziswa kuita autoregressive spiking language modhi. Nzira yacho inodzidzisa pazvivakashure zvinogadzirwa nemuenzaniso wemudzidzi, uku ichiwedzera zvidzoreso zvinotarisirwa kuchengetedza kuchinjika kwakadzikama.
Iro bepa rinogadzirisa dambudziko chairo mukushandura akajairwa mutauro modhi kuita spiking neural network vadzidzi. Mukuseta kwakatsanangurwa nevanyori, mudzidzisi akadzidziswa neural network anotarisisa mudzidzi ari spiking kuburikidza neruzivo distillation. Nzira dziripo dzekufambisa dzinodzidzisa pazvivakamberi zvakamisikidzwa zvakatorwa kubva kucorpus, nepo autoregressive inference inoshandisa prefixes inogadzirwa nemodhi yacho pachayo. Vanyori vanoratidza musiyano iwoyo se prefix-source mismatch.
Maererano nebepa, kusawirirana kune maitiro maviri. Inogona kuburitsa inobuditsa-mutemo kusawirirana pakati peiyo artificial mudzidzisi uye mudzidzi, uye inogona kukonzera kuti mukati memudzidzi spiking dynamics ikukure kana ichigadzira ega anozvigadzira prefixes pane kuenderana corpus prefixes. On-policy distillation inoratidzwa senzira yekufumura kudzidziswa kune izvo zvakazvigadzira prefixes uye kuenderera mberi netarisiro yevadzidzisi pasi pemamiriro ari pedyo nekushandisa autoregressive.
Vanyori vanotanga vaongorora mudzidzisi-chete yakazara-KL vhezheni inonzi Vanilla OPD muyedzo inodzorwa yekushushikana. Vanoshuma kuti nzira iyi inogona kusangana nekunonoka kuburitsa-mhinduro kudonha. Muakaundi yebepa, kutsvaga ikoko kunoreva kuti kungowedzera kuratidzwa kune anozvigadzira ega prefixes hakuvimbisi kugadzirisa kwakagadzikana. Mhedzisiro yacho inokurudzira chimiro chinosanganisa kudzidza-kwemitemo nezvimwe zvipingamupinyi zvekuti mudzidzi angafamba kusvika papi panguva yekudzidziswa.
SpikeOPD inoshandisa zvinhu zvitatu zvinotsanangurwa mune abstract. Yakazara-KL gadziriso yemudzidzisi inoitirwa kudzikisa kuburitsa-mutemo kusawirirana. Inofananidzwa-prefix policy anchoring constrains policy kusimuka kubva kune yakaomeswa referensi spiking modhi pane imwechete prefixes. Layerwise spike regularization inodzikamisa kutsauka mumitengo yekupfura panguva ye-policy adaptation. Pamwe chete, nzira idzi dzakagadzirirwa kuchengetedza kugadzikana kwekuburitsa uku ichibvumira mudzidzi kudzidza kubva kune yaanogadzira.
Vanyori vanorondedzera ongororo pazvikero zvitatu zvemuenzaniso zvakanyorwa 0.125B, 0.35B uye 1.3B. Zvinei kune inoenderana ruzivo-distillation spiking modhi, SpikeOPD inovandudza avhareji yechokwadi ne0.8, 1.7 uye 2.9 mapoinzi, zvichiteerana. Tsime rinoti zviwanikwa izvi zvinowanwa uchichengetedza mamodheru 'sparse-compute profiles. Iyo abstract haipe iyo iri pasi pebasa runyorwa, dhataseti, zvibodzwa zvekutanga, kusava nechokwadi kwehuwandu kana akadzama hardware mamiriro.
Nei zvichikosha
Mitauro yeSpiking iri kuongororwa senzira yekuwedzera simba rekugadzirisa mutauro kuburikidza nekombuta shoma, inofambiswa nezviitiko. Mhedzisiro dzakashumwa dzinoratidza kuti kugadzirisa kusawirirana pakati pekudzidziswa kwekuisa uye hunhu hwekufungidzira kunogona kuvandudza huroyi pasina kurasa sparse-compute maitiro anokurudzira spiking modhi.
Iko kukosha kwekushanda kwebasa kunosungirirwa kune tradeoff mumamodheru emitauro. Tsime rinotsanangura mashoma encoding uye chiitiko-inotyairwa computation senzira inogoneka kune simba-inoshanda mutauro modhi, asi zvakare inotaura kuti kudzidzisa inokwanisa spiking mitauro yemhando kubva kutanga kunoramba kwakaoma. Yakabudirira artificial-neural-network-to-spiking migration nzira inogona saka kubatsira kana ikavandudza kugona pasina kubvisa computational zvimiro zvinoita kuti spiking masisitimu anakidze.
Mupiro wepakati wepepa hausi kungotaura chokwadi. Inotarisa pamukaha wemaitiro pakati pemadzidzisirwo emhando yemutauro uye kuti inoburitsa sei zvinyorwa. Nekuti autoregressive masisitimu anomira pane zvaakamboburitsa, nzira yekudzidzira iyo inomiririra yeanozvigadzira ega prefixes inogona kugadzirisa kutadza maitiro ayo fixed-corpus-prefix kudzidziswa isingafumure. Izvo zvinotarisirwa kugadzikana zvinodzora zvine chekuita nevatsvaguri vanoedza kuita kuti vadzidzi ve spiking vazvibate zvakavimbika munguva yechizvarwa chakawedzerwa.
Izvo zvakashumwa zvinowedzera nechiyero chemuenzaniso wakanyorwa, kubva pa 0.8 points pa 0.125B kusvika 2.9 points pa 1.3B. Kana ikagadzirwazve, patani iyoyo yaizoita kuti nzira yacho ive yakakosha kune mamodheru makuru akatamiswa, kunyangwe kwakabva kusina tsananguro yekuti sei mibairo ichiwedzera kana kuti inoenderera mberi kupfuura zviyero zvitatu zvakaongororwa. Mhedzisiro yacho saka humbowo hwemagadzirirwo akaedzwa, kwete ratidziro yakajairika yekuti ese mamodheru emitauro inovandudza nenzira imwechete.
Kuchengetedza sparse-compute profiles zvine basa nekuti kuvandudzwa kwechokwadi kwaida kusiya maitiro mashoma kunonetesa chikonzero chemaitiro. Tsime racho rinotaura zvakajeka kuti maprofile akachengetedzwa, asi haashandure chirevo ichocho kuita simba, latency, ndangariro kana zviyero zvekushandisa-mutengo. Vaverengi vanofanirwa kusiyanisa pepa rekuvaka-komputa kudai kubva kune yakasimbiswa kudzikiswa kwekushandiswa kwemagetsi kana mutengo wekutumira.
Basa racho rinogona kupawo nzira yekudzidzisa yekirasi yakafara yeautoregressive mudzidzi modhi, asi sosi inongosimbisa iyo yakarongwa chimiro mune yespiking-mutauro-modhi yekumisikidza yakatsanangurwa. Iyo hairatidze sisitimu yekugadzira, chigadzirwa chakatarisana nemushandisi, kutumirwa, kana yakazvimirira yakaongororwa kubatsirwa kune vatengi. Kukosha kwayo kwakaringana semhedzisiro yehunyanzvi iyo vaongorori vanogona kuongorora, kubereka uye kupikisa.
Interactive Mechanism: Iyo Inonyatsoshanda
Ongorora ari pasi tekinoroji kuseri kwekusimudzira uku uchipindirana.
Which component of an AI application is the machine-learning model itself?
Zvekutarisa zvinotevera
Iro bepa ndeye arXiv vhezheni-imwe preprint, uye sosi haigadziri yakazvimirira kudzokorora, ongororo yevezera, kutumirwa kwepasirese, kana kuyerwa kwekuchengetedza simba. Basa rekutevera rinofanirwa kuyedza kana izvo zvakaziviswa zvakabata mabasa akafara, dhatasethi, Hardware uye inference mamiriro.
Mubvunzo wekutanga ndeye reproducibility. Kwanobva kunozivisa bepa searXiv:2608.27857, vhezheni yekutanga, yakatumirwa Aug. 28, 2026. Haitauri yakazvimirira kudzokorora kana mamiriro ekuongorora-vezera. Ongororo dzekutevera dzinofanirwa kuona kuvandudzwa kwechokwadi kunoenderana neino enderana ruzivo-distillation nheyo uye kujekesa kana kuenzanisa kunoshandisa zvakafanana data, mabhajeti ekudzidzisa uye maitiro ekuongorora.
Iyo abstract haitauri mabasa kana dhatabheti rinoshandiswa kuverenga avhareji yechokwadi. Kusiiwa ikoko kunosiya mibvunzo yakakosha isina kugadziriswa: kana zviwanikwa zvakanangana mune mamwe masimba, angave achichinjisa munzvimbo dzese, uye kana avhareji yechokwadi inovanza kudzokororwa pamabasa ega. Kuzivisa kwakadzama kwe-per-task mhedzisiro, matanho ekuvimba uye nyaya dzekukundikana kwaizoita kuti kudai kuve nyore kuongorora.
Chirevo chekugadzikana kwepepa chinobvumirawo kuongororwa kwekushushikana. Vanilla OPD inonzi yakatambura kunonoka kuburitsa-mhinduro kudonha muyedzo inodzorwa, nepo SpikeOPD yakagadzirirwa kuchengetedza kugadzikana kwekuburitsa. Basa remangwana rinofanirwa kuongorora zvizvarwa zvenguva refu, mamiriro ekuenzanisa akasiyana, maprefixes akareba ega uye marongero ayo mudzidzi anosiyana zvakanyanya kubva kumudzidzisi. Kwakabva hakuratidze kuti nzira yacho yakasimba sei kunze kwezviyedzo zvakashumwa.
Kushanda kwesimba kunoramba kuri mubvunzo wakavhurika. Tsime rinoti spiking inopa nzira yekusimba-inoshanda mutauro modhi uye inoti SpikeOPD inochengetedza sparse-compute profiles, asi haipe yakananga simba, latency, throughput kana hardware zviyero muzvinyorwa zvinopihwa. Kuongorora kunoshanda kwaizoda kuyera izvo mhedzisiro pane yakakodzera Hardware pane kuti iite kubva kune sparsity chete.
Pakupedzisira, sosi haigadziri kugadzirira kwekutumira. Zvisingazivikanwe zvinosanganisira mutengo wekudzidziswa we-on-policy distillation, mutengo wakawedzerwa wekugadziriswa kwevadzidzisi uye kuita nguva dzose, zvinodikanwa zvekuyeuka zvekuchengetedza marongero ereferenzi, uye maitiro emaitiro pazvikero zvakakura. Izvo zvinhu zvinosarudza kana izvo zvakashumwa zvakawanikwa zvichishandura kuita mukana unobatsira kune chaiwo masisitimu.