Aikace-aikacen sarrafa bayanai na ainihi na buƙatar ƙayyadaddun tsarin ƙididdiga, ƙananan latency, ƙananan ƙarfi. Tare da ikon sarrafa abubuwan da ke haifar da aukuwa, ƙarin ƙarfe-oxide-semiconductor hybrid memristive neuromorphic architectures suna ba da ingantaccen tushe na kayan aiki don irin waɗannan ayyuka. Don nuna cikakken yuwuwar irin waɗannan tsarin, muna ba da shawara kuma mu gwada gwaji don nuna cikakken bayani na sarrafa firikwensin don aikace-aikacen keɓance abu na ainihi. Zane wahayi daga sito owl neuroanatomy, mun ɓullo da wani bioinspirationd, aukuwa-kore tsarin gano wuri abu wanda ya haɗu da zamani na zamani piezoelectric micromechanical transducer transducer tare da lissafi na tushen neuromorphic resistive memory. Muna nuna ma'auni na tsarin ƙirƙira wanda ya haɗa da na'ura mai gano juriya na daidaituwar daidaituwa na tushen ƙwaƙwalwar ajiya, jinkirin layin kewayawa, da cikakkiyar transducer ultrasonic na musamman. Muna amfani da waɗannan sakamakon gwaji don daidaita simulations a matakin tsarin. Ana amfani da waɗannan simintin don ƙididdige ƙudurin kusurwa da ingancin kuzarin samfurin gano wuri. Sakamakon ya nuna cewa tsarinmu na iya zama umarni da yawa na girma mafi ƙarfi fiye da masu sarrafa microcontrollers waɗanda ke yin aiki iri ɗaya.
Muna shiga wani zamani na sarrafa kwamfuta a ko'ina inda adadin na'urori da tsarin da aka tura ke karuwa sosai don taimaka mana a rayuwarmu ta yau da kullun. Ana sa ran waɗannan tsarin za su ci gaba da gudana, suna cinye ƙarancin ƙarfi kamar yadda zai yiwu yayin da suke koyon fassarar bayanan da suke tattarawa daga na'urori masu auna firikwensin a cikin ainihin lokaci da kuma samar da fitarwa na binary sakamakon rarrabuwa ko ayyukan tantancewa. Ɗaya daga cikin mahimman matakan da ake buƙata don cimma wannan burin shine ciro bayanai masu amfani da ƙaƙƙarfan bayanai daga hayaniya da sau da yawa marasa cikakkun bayanai. Hannun aikin injiniya na al'ada yawanci samfurin siginar firikwensin a koyaushe kuma mai girma, yana samar da adadi mai yawa na bayanai ko da in babu bayanai masu amfani. Bugu da kari, waɗannan hanyoyin suna amfani da hadaddun dabarun sarrafa siginar dijital don aiwatar da bayanan shigarwa (sau da yawa masu hayaniya). Madadin haka, ilmin halitta yana ba da madadin mafita don sarrafa bayanai masu hayaniya ta amfani da ingantaccen kuzari, asynchronous, hanyoyin tafiyar da aukuwa (spikes)2,3. Ƙididdigar Neuromorphic yana ɗaukar wahayi daga tsarin ilimin halitta don rage farashin ƙididdiga dangane da makamashi da buƙatun ƙwaƙwalwar ajiya idan aka kwatanta da hanyoyin sarrafa siginar gargajiya4,5,6. Kwanan nan, sabbin tsarin tushen kwakwalwa na gabaɗaya waɗanda ke aiwatar da hanyoyin sadarwa na jijiya (TrueNorth7, BrainScaleS8, DYNAP-SE9, Loihi10, Spinnaker11) an nuna su. Waɗannan na'urori masu sarrafawa suna ba da ƙarancin ƙarfi, ƙarancin latency mafita don koyan inji da ƙirar kewayen cortical. Don cikakken amfani da ƙarfin kuzarinsu, waɗannan na'urori masu sarrafa neuromorphic dole ne a haɗa su kai tsaye zuwa na'urori masu auna firikwensin taron12,13. Koyaya, a yau akwai ƴan na'urorin taɓawa waɗanda ke ba da bayanan da ke gudana kai tsaye. Shahararrun misalan su ne firikwensin gani mai ƙarfi (DVS) don aikace-aikacen hangen nesa kamar bin diddigi da gano motsi14,15,16,17 da silicon cochlea18 da neuromorphic auditory firikwensin (NAS)19 don sarrafa siginar sauraro, firikwensin olfactory20 da misalai masu yawa21,22 na taɓawa. . na'urori masu auna sigina.
A cikin wannan takarda, mun gabatar da sabon tsarin sarrafa ji na abin da aka ɓullo da shi wanda aka yi amfani da shi don gano abubuwa. Anan, a karo na farko, mun bayyana tsarin ƙarshe zuwa ƙarshen don ƙaddamar da abu da aka samu ta hanyar haɗawa da fasahar zamani na piezoelectric micromachined ultrasonic transducer (pMUT) tare da jadawali na lissafi dangane da ƙwaƙwalwar ƙwaƙwalwar ƙwaƙwalwar ƙwaƙwalwar ƙwaƙwalwa ta neuromorphic (RRAM). Gine-ginen ƙirar ƙwaƙwalwar ajiya ta amfani da RRAM mafita ce mai ban sha'awa don rage yawan amfani da wutar lantarki23,24,25,26,27,28,29. Rashin rashin daidaituwarsu na asali-ba buƙatar amfani da wutar lantarki mai aiki don adanawa ko sabunta bayanai ba-daidai ne tare da asynchronous, yanayin korar abubuwan da ke haifar da ƙididdiga na neuromorphic, yana haifar da kusan-ba amfani da wutar lantarki lokacin da tsarin ba shi da aiki. Piezoelectric micromachined ultrasonic transducers (pMUTs) ba su da tsada, ƙaramin siliki na tushen ultrasonic transducers masu iya aiki azaman masu watsawa da masu karɓa30,31,32,33,34. Don aiwatar da siginar da aka samu ta hanyar ginanniyar firikwensin, mun zana wahayi daga barn owl neuroanatomy35,36,37. Mujiya sito Tyto alba sananne ne saboda iyawar farautarsa na dare saboda ingantaccen tsarin ganowa. Don ƙididdige wurin da aka fara ganima, tsarin ƙayyadaddun mujiya na sito yana ɓoye lokacin tashi (ToF) lokacin da raƙuman sauti daga ganima ya isa kowane kunnen mujiya ko masu karɓar sauti. Idan aka yi la'akari da nisa tsakanin kunnuwa, bambanci tsakanin ma'aunin ToF guda biyu (Bambancin Lokacin Interaural, ITD) yana ba da damar ƙididdige matsayin azimuth na manufa. Ko da yake tsarin ilimin halitta bai dace da warware lissafin algebra ba, za su iya magance matsalolin gida da kyau sosai. Tsarin jijiya na mujiya yana amfani da saitin mai gano daidaituwa (CD) 35 neurons (watau neurons masu iya gano alaƙar ɗan lokaci tsakanin spikes waɗanda ke yaduwa zuwa ƙasa zuwa ƙarshen tashin hankali)38,39 an tsara su cikin jadawali na lissafi don magance matsalolin sakawa.
Binciken da ya gabata ya nuna cewa kayan aikin kayan aikin ƙarfe-oxide-semiconductor (CMOS) da kayan aikin neuromorphic na tushen RRAM wanda aka yi wahayi zuwa ga ƙananan colliculus (“auditory cortex”) na mujiya sito wata ingantacciyar hanya ce don ƙididdige matsayi ta amfani da ITD13, 40, 41, 42, 43, 44, 45, 46. Duk da haka, yuwuwar cikakken tsarin tsarin neuromorphic wanda ke danganta alamomin saurare zuwa jadawali na lissafin neuromorphic har yanzu ba a nuna su ba. Babban matsalar ita ce bambancin da'irar analog na CMOS, wanda ke shafar daidaiton gano wasa. Kwanan nan, madadin aiwatar da ƙididdiga na ITD47 an nuna. A cikin wannan takarda, muna ba da shawarar yin amfani da ikon RRAM don canza ƙimar gudanarwa ta hanyar da ba ta da ƙarfi don magance sauye-sauye a cikin da'irori na analog. Mun aiwatar da tsarin gwaji wanda ya ƙunshi membrane mai watsa pMUT guda ɗaya wanda ke aiki akan mitar 111.9 kHz, pMUT biyu masu karɓar membranes (sensors) na simulating kunnuwa na mujiya, da ɗaya. Mun gwada tsarin gano pMUT da jadawali na tushen RRAM don gwada tsarin gurɓatar mu da kimanta ƙudurinsa na kusurwa.
Muna kwatanta hanyarmu tare da aiwatarwa na dijital akan microcontroller wanda ke yin aiki iri ɗaya ta hanyar amfani da ƙirar katako na al'ada ko hanyoyin neuromorphic, da kuma tsarin tsararrun kofa (FPGA) don kimanta ITD da aka gabatar a cikin tunani. 47. Wannan kwatancen yana ba da haske ga ingantaccen ƙarfin ƙarfin tsarin RRAM na tushen analog neuromorphic tsarin.
Ana iya samun ɗaya daga cikin misalan mafi ɗaukar hankali na ingantaccen kuma ingantaccen tsarin gano abu a cikin sito owl35,37,48. Da magariba da wayewar gari, mujiya sito (Tyto Alba) da farko ta dogara ne da sauraren ra'ayi, da ƙwaƙƙwaran neman ƙaramin ganima kamar voles ko mice. Waɗannan ƙwararrun masu saurare na iya gano siginar ji daga ganima tare da daidaito mai ban mamaki (kimanin 2°)35, kamar yadda aka nuna a hoto. 1a. Owls na Barn suna fahimtar wurin tushen sauti a cikin jirgin azimuth (a kwance) daga bambancin lokacin shigowar jirgi (ITD) daga tushen sauti zuwa kunnuwa biyu. Jeffress49,50 ne ya gabatar da tsarin lissafin ITD wanda ya dogara da lissafi na jijiya kuma yana buƙatar mahimman abubuwa guda biyu: axon, fiber jijiya na neuron wanda ke aiki azaman layin jinkiri, da tsararrun ƙwayoyin gano ƙwayoyin cuta waɗanda aka tsara cikin tsarin lissafi. jadawali kamar yadda aka nuna a hoto na 1b. Sautin yana kaiwa kunne tare da jinkirin dogaro da azimuth (ITD). Sa'an nan kuma ana juyar da sautin zuwa tsarin karu a kowace kunne. Axon na kunnuwan hagu da na dama suna aiki azaman layin jinkiri kuma suna haɗuwa a kan ƙwayoyin CD. A ka'ida, neuron guda ɗaya kawai a cikin tsararrun na'urorin da suka dace zasu karɓi shigarwa a lokaci guda (inda jinkirin ya soke daidai) kuma zai yi wuta da yawa (kwayoyin maƙwabta su ma za su yi wuta, amma a ƙananan mitar). Kunna wasu neurons suna ɓoye matsayin manufa a sararin samaniya ba tare da ƙara canza ITD zuwa kusurwoyi ba. An taƙaita wannan ra'ayi a cikin Hoto 1c: alal misali, idan sautin yana fitowa daga gefen dama lokacin da siginar shigarwa daga kunnen dama ya yi tafiya mai tsawo fiye da hanyar daga kunnen hagu, yana ramawa ga adadin ITDs, misali. lokacin da neuron 2 ya dace. A wasu kalmomi, kowane CD yana amsawa ga wani ITD (wanda kuma aka sani da jinkiri mafi kyau) saboda jinkirin axonal. Don haka, kwakwalwa tana jujjuya bayanan lokaci zuwa bayanin sarari. An sami shaidar ilimin halitta na wannan tsarin37,51. Neuron macronucleus masu kulle-kulle suna adana bayanai na ɗan lokaci game da sautuna masu shigowa: kamar yadda sunansu ke nunawa, suna harbi a wasu matakan sigina. Za'a iya samun na'urar gano ma'anar daidaituwa na ƙirar Jeffress a cikin laminar core. Suna karɓar bayanai daga ƙananan ƙwayoyin cuta na macronuclear, waɗanda axon suna aiki azaman layin jinkiri. Adadin jinkirin da aka bayar ta hanyar jinkirin layin za'a iya bayyana shi ta tsawon tsayin axon, da kuma wani nau'in myelination wanda ke canza saurin tafiyarwa. An yi wahayi zuwa ga tsarin sauraren mujiya na sito, mun haɓaka tsarin biomimetic don gano abubuwa. Kunnuwa biyu suna wakiltar masu karɓar pMUT guda biyu. Tushen sauti shine mai watsa pMUT da ke tsakanin su (Fig. 1a), kuma jadawali yana samuwa ta hanyar grid na CD na tushen RRAM (Fig. 1b, kore), yana taka rawar CD neurons waɗanda aka jinkirta shigar da su. ta hanyar kewayawa, layin jinkiri (blue) suna aiki kamar axon a cikin takwarorinsu na nazarin halittu. Tsarin azanci da aka tsara ya bambanta a mitar aiki da na mujiya, wanda tsarin sauraronsa ke aiki a cikin kewayon 1-8 kHz, amma ana amfani da firikwensin pMUT da ke aiki a kusan 117 kHz a cikin wannan aikin. Ana la'akari da zaɓin mai jujjuyawar ultrasonic bisa ga ka'idodin fasaha da haɓakawa. Na farko, iyakance bandwidth karɓar zuwa mitar guda ɗaya da kyau yana inganta daidaiton aunawa kuma yana sauƙaƙa matakin aiwatarwa. Bugu da ƙari, yin aiki a cikin duban dan tayi yana da fa'ida cewa bugun jini da aka fitar ba a ji ba, don haka kada ku dame mutane, tunda kewayon sauraron su shine ~ 20-20 kHz.
mujiya sito yana karɓar raƙuman sauti daga wani manufa, a cikin wannan yanayin yana motsa ganima. Lokacin tashi (ToF) na motsin sauti ya bambanta ga kowane kunne (sai dai idan ganima yana gaban mujiya kai tsaye). Layin mai dige-dige yana nuna hanyar da raƙuman sauti ke bi don isa kunnuwan mujiya. Ana iya gano ganima daidai a cikin jirgin sama a kwance bisa tsayin daka tsakanin hanyoyin sauti guda biyu da bambancin lokacin tsaka-tsakin lokaci (ITD) (hoton hagu da aka yi wahayi ta hanyar ref. 74, haƙƙin mallaka 2002, Society for Neuroscience). A cikin tsarin mu, mai watsa pMUT (duhu mai duhu) yana haifar da raƙuman sauti waɗanda ke billa kan manufa. Ana karɓar raƙuman duban dan tayi ta masu karɓar pMUT guda biyu (kore mai haske) kuma ana sarrafa su ta hanyar injin neuromorphic (dama). b Wani samfurin lissafi na ITD (Jeffress) wanda ke kwatanta yadda sautin da ke shiga kunnuwan sito na mujiya aka fara sanya su azaman ƙulle-ƙulle a cikin babban tsakiya (NM) sannan ta yin amfani da grid ɗin da aka tsara ta geometrically na ma'aunin gano ƙwayoyin jijiya a cikin tsakiya na lamellar. Gudanarwa (Netherland) (hagu). Misalin jadawali na lissafin neuroITD wanda ya haɗa layin jinkiri da ƙwayoyin gano ƙwayoyin cuta, ana iya ƙirƙira tsarin biosensor na mujiya ta amfani da da'irar neuromorphic na tushen RRAM (dama). c Tsarin tsari na babban tsarin Jeffress, saboda bambanci a cikin ToF, kunnuwa biyu suna karɓar sautin sauti a lokuta daban-daban kuma suna aika axon daga ƙarshen biyu zuwa mai ganowa. Axon wani yanki ne na jerin abubuwan gano daidaituwa (CD) neurons, kowannensu yana ba da amsa zaɓaɓɓu ga abubuwan da ke da alaƙa da lokaci mai ƙarfi. Sakamakon haka, CD kawai waɗanda abubuwan shigar su suka zo tare da mafi ƙarancin lokacin bambance-bambancen suna da matuƙar farin ciki (ITD daidai aka biya). CD ɗin za ta ɓoye matsayin kusurwar manufa.
Piezoelectric micromechanical ultrasonic transducers ne scalable ultrasonic transducers da za a iya hadedde tare da ci-gaba fasahar CMOS31,32,33,52 kuma suna da ƙananan ƙarfin lantarki na farko da kuma amfani da wutar lantarki fiye da na gargajiya volumetric transducers53. A cikin aikinmu, diamita na membrane shine 880 µm, kuma ana rarraba mitar resonant a cikin kewayon 110-117 kHz (Fig. 2a, duba Hanyoyi don cikakkun bayanai). A cikin nau'in na'urorin gwaji guda goma, matsakaicin ma'aunin inganci ya kasance kusan 50 (misali 31). Fasahar ta kai girman masana'antu kuma ba ta da kwarjini a kowane iri. Haɗa bayanai daga fina-finai daban-daban na pMUT sanannen fasaha ne, kuma ana iya samun bayanin kusurwa daga pMUTs ta amfani da, misali, fasahar beamforming31,54. Koyaya, sarrafa siginar da ake buƙata don cire bayanan kusurwa bai dace da ƙananan ma'aunin wuta ba. Tsarin da aka tsara ya haɗu da bayanan neuromorphic da ke sarrafa pMUT tare da jadawali na tushen neuromorphic na RRAM wanda aka yi wahayi ta hanyar ƙirar Jeffress (Hoto 2c), yana ba da madadin ingantaccen makamashi da ingantaccen kayan masarufi. Mun yi gwaji inda aka sanya firikwensin pMUT guda biyu kusan 10 cm tsakanin su don yin amfani da sautunan ToF daban-daban da maɓallan biyu masu karɓa suka karɓa. Ɗayan pMUT mai aiki azaman mai watsawa yana zaune tsakanin masu karɓa. Manufar ita ce farantin PVC 12 cm fadi, wanda yake a nesa D a gaban na'urar pMUT (Fig. 2b). Mai karɓa yana rikodin sautin da aka nuna daga abu kuma yana mayar da martani gwargwadon iko yayin tafiyar motsin sauti. Maimaita gwajin ta canza matsayin abu, an ƙaddara ta nisa D da kusurwa θ. Ƙaddamar da hanyar haɗi. 55, muna ba da shawarar pre-aiki na neuromorphic na pMUT raw sigina don canza raƙuman raƙuman ruwa zuwa kololuwa don shigar da jadawali na lissafin neuromorphic. Ana fitar da ToF daidai da girman girman kololuwa daga kowane tashoshi biyu kuma an sanya shi azaman ainihin lokacin kololuwar mutum ɗaya. A kan fig. 2c yana nuna kewayawar da ake buƙata don yin mu'amala da firikwensin pMUT tare da jadawali na tushen RRAM: ga kowane ɗayan masu karɓar pMUT guda biyu, siginar siginar da aka tace ta band-pass don santsi, gyarawa, sannan ta wuce zuwa ga mai haɗawa da leaky cikin yanayin cin nasara. ƙofa mai ƙarfi (Fig. 2d) yana haifar da taron fitarwa (karu) da harbe-harbe (LIF) neuron: lokacin ƙaruwar fitarwa yana ɓoye lokacin jirgin da aka gano. An daidaita madaidaicin LIF akan martanin pMUT, don haka rage yawan pMUT daga na'ura zuwa na'ura. Tare da wannan hanyar, maimakon adana dukkan igiyoyin sauti a cikin ƙwaƙwalwar ajiya da sarrafa shi daga baya, kawai muna samar da kololuwa daidai da ToF na igiyoyin sauti, wanda ke samar da shigarwar zuwa jadawali mai ƙididdigewa. Ana aika spikes kai tsaye zuwa layin jinkiri kuma an daidaita su tare da samfuran gano wasa a cikin jadawalin lissafin neuromorphic. Domin ana aika su zuwa ƙofofin transistor, ba a buƙatar ƙarin da'irar ƙarawa (duba Ƙarin Hoto 4 don cikakkun bayanai). Don kimanta daidaitattun angular angular da aka bayar ta pMUT da kuma hanyar sarrafa siginar da aka tsara, mun auna ITD (wato, bambancin lokaci tsakanin kololuwar al'amuran da masu karɓa biyu suka samar) yayin da nisa da kusurwar abu suka bambanta. An canza nazarin ITD zuwa kusurwoyi (duba Hanyoyi) kuma an tsara shi a kan matsayi na abu: rashin tabbas a cikin ma'auni na ITD ya karu da nisa da kusurwa zuwa abu (Fig. 2e, f). Babban matsalar ita ce ƙimar kololuwa zuwa amo (PNR) a cikin martanin pMUT. Mafi nisa abu, ƙananan siginar sauti, don haka rage PNR (Fig. 2f, layin kore). Ragewa a cikin PNR yana haifar da karuwa a cikin rashin tabbas a cikin ƙididdigar ITD, wanda ya haifar da karuwa a cikin daidaituwa (Fig. 2f, blue line). Don wani abu a nesa na 50 cm daga mai watsawa, daidaiton kusurwar tsarin shine kusan 10 °. Wannan iyakancewar da aka sanya ta halayen firikwensin za a iya inganta shi. Misali, ana iya ƙara matsa lamba da emitter ya aika, ta haka zai ƙara ƙarfin lantarki da ke motsa membrane pMUT. Wani bayani don haɓaka siginar da aka watsa shi ne haɗa masu watsawa da yawa 56. Waɗannan mafita za su kara yawan adadin ganowa a kashe kuɗin da ake samu na makamashi. Ana iya ƙara ƙarin haɓakawa akan ɓangaren karɓa. Za a iya rage girman hayaniyar mai karɓar pMUT ta hanyar inganta haɗin kai tsakanin pMUT da amplifier mataki na farko, wanda a halin yanzu ana yin shi tare da haɗin waya da igiyoyi RJ45.
Hoton kristal pMUT tare da membranes 880 µm da aka haɗa a cikin farar 1.5 mm. b Zane na saitin aunawa. Maƙasudin yana samuwa a matsayi na azimuth θ kuma a nesa D. Mai watsawa na pMUT yana haifar da siginar 117.6 kHz wanda ya tashi daga maƙasudin kuma ya kai masu karɓar pMUT guda biyu tare da lokaci-lokaci daban-daban (ToF). Wannan bambance-bambancen, wanda aka ayyana azaman bambancin lokacin jita-jita (ITD), yana ɓoye matsayin abu kuma ana iya ƙididdige martanin kololuwar na'urorin masu karɓa guda biyu. c Tsare-tsare na matakan aiwatarwa don canza siginar pMUT danye zuwa jeri mai karu (watau shigarwa zuwa jadawali na lissafin neuromorphic). An ƙirƙira kuma an gwada na'urorin firikwensin pMUT da zane-zanen lissafi na neuromorphic, kuma aikin pre-processing na neuromorphic ya dogara ne akan simintin software. Amsar da membrane pMUT akan karɓar sigina da kuma canza shi zuwa yanki mai karu. e Gwajin gano daidaitaccen kusurwa azaman aikin kusurwar abu (Θ) da nisa (D) zuwa abin da aka yi niyya. Hanyar hakar ITD tana buƙatar ƙaramin ƙuduri na kusurwa na kusan 4°C. f Daidaiton kusurwa (layin shuɗi) da daidaitaccen rabo-zuwa amo (layin kore) tare da nisan abu don Θ = 0.
Ƙwaƙwalwar ajiya mai juriya tana adana bayanai a cikin yanayin da ba mara maras tabbas ba. Babban ka'idar hanyar ita ce gyare-gyaren abu a matakin atomic yana haifar da canji a cikin ƙarfin lantarki57. Anan muna amfani da ƙwaƙwalwar juzu'i na tushen oxide wanda ya ƙunshi 5nm Layer na hafnium dioxide sandwiched tsakanin sama da ƙasa titanium da titanium nitride electrodes. Za'a iya canza yanayin tafiyar da na'urorin RRAM ta hanyar amfani da sigar halin yanzu/ƙarar wutar lantarki wanda ke ƙirƙira ko karya filaments na guraben iskar oxygen tsakanin wayoyin lantarki. Mun haɗa irin waɗannan na'urori58 a cikin daidaitaccen tsari na 130 nm na CMOS don ƙirƙirar da'irar da za'a iya sake daidaitawa na neuromorphic aiwatar da mai gano daidaituwa da da'irar layin jinkiri (Fig. 3a). Halin rashin daidaituwa da na'urar na'urar, haɗe tare da yanayin da'irar da'irar neuromorphic, yana rage yawan amfani da wutar lantarki. Da'irar tana da aikin kunnawa/kashe nan take: yana aiki nan da nan bayan an kunna shi, yana ba da damar kashe wutar gabaɗaya lokacin da kewaye ba ta da aiki. Ana nuna babban tubalan ginin tsarin da aka tsara a cikin fig. 3 b. Ya ƙunshi tsarin N layi ɗaya-resistor single-transistor (1T1R) wanda ke ɓoye ma'aunin synaptic daga abin da ake ɗaukar igiyoyi masu nauyi, allura a cikin synapse na gama gari na mai haɗawa biyu (DPI) 59, kuma a ƙarshe an allura a cikin synapse tare da haɗin kai yabo. kunna (LIF) neuron 60 (duba Hanyoyi don cikakkun bayanai). Ana amfani da matakan shigarwa zuwa ƙofar tsarin 1T1R a cikin nau'i na nau'i na nau'i na nau'i na ƙarfin lantarki tare da tsawon lokaci akan tsari na daruruwan nanoseconds. Ƙwaƙwalwar ƙwaƙwalwar ajiya za a iya sanyawa a cikin babban yanayin gudanarwa (HCS) ta hanyar yin amfani da ma'anar tabbatacce ta waje zuwa Vtop lokacin da Vbottom ya kasance ƙasa, da kuma sake saitawa zuwa ƙasa maras kyau (LCS) ta amfani da ingantaccen ƙarfin lantarki zuwa Vbottom lokacin da Vtop ke ƙasa. Ana iya sarrafa matsakaicin ƙimar HCS ta iyakance shirye-shiryen halin yanzu (cirewa) na SET (ICC) ta hanyar wutar lantarki ta tushen ƙofar jerin transistor (Fig. 3c). Ayyukan RRAM a cikin da'irar suna da ninki biyu: suna jagora da auna bugun bugun shigar.
Hoton microscope na lantarki (SEM) na na'urar HfO2 1T1R RRAM shuɗi wanda aka haɗa a cikin fasahar CMOS nm 130 tare da transistor masu zaɓi (fadi 650 nm) a kore. b Tubalan ginin asali na tsarin tsarin neuromorphic. Matsakaicin wutar lantarki na shigarwa (kololuwa) Vin0 da Vin1 suna cinye Iweight na yanzu, wanda yayi daidai da jahohin gudanarwa G0 da G1 na tsarin 1T1R. Ana shigar da wannan halin yanzu a cikin synapses na DPI kuma yana farantawa LIF neurons. An shigar da RRAM G0 da G1 a cikin HCS da LCS bi da bi. c Aiki na jimlar gudanarwa na rukuni na na'urorin 16K RRAM a matsayin aikin ICC na yanzu, wanda ke sarrafa matakin gudanarwa yadda ya kamata. d Ma'aunin kewayawa a cikin (a) yana nuna cewa G1 (a cikin LCS) yana toshe shigarwa daga Vin1 (kore), kuma haƙiƙa ƙarfin wutar lantarki na neuron yana amsawa kawai ga shigarwar shuɗi daga Vin0. RRAM yana ƙayyadad da haɗin haɗin da ke cikin kewaye yadda ya kamata. e Aunawar da'irar a cikin (b) yana nuna tasirin ƙimar gudanarwar G0 akan ƙarfin lantarki na membrane Vmem bayan amfani da bugun bugun jini Vin0. Yawancin tafiyarwa, mafi ƙarfin amsa: don haka, na'urar RRAM tana aiwatar da ma'aunin haɗin I/O. An yi ma'auni akan kewayawa kuma suna nuna ayyuka biyu na RRAM, kewayawa da ma'auni na bugun bugun jini.
Na farko, tun da akwai manyan jahohin gudanarwa guda biyu (HCS da LCS), RRAMs na iya toshewa ko rasa abubuwan shigarwa lokacin da suke cikin jihohin LCS ko HCS, bi da bi. Sakamakon haka, RRAM yana ƙayyadad da haɗin haɗin da ke cikin kewaye. Wannan shine tushen samun damar sake fasalin gine-gine. Don nuna wannan, za mu bayyana ƙayyadaddun aiwatar da aikin da'ira a cikin siffa 3b. RRAM mai dacewa da G0 an tsara shi cikin HCS, kuma RRAM G1 na biyu an tsara shi cikin LCS. Ana amfani da ƙwanƙolin shigarwa akan duka Vin0 da Vin1. An yi nazari akan tasirin jeri biyu na bugun jini na shigarwa a cikin jijiyoyi masu fitarwa ta hanyar tattara wutar lantarki ta neuron membrane da siginar fitarwa ta amfani da oscilloscope. Gwajin ya yi nasara lokacin da na'urar HCS (G0) kawai aka haɗa da bugun jini na neuron don tada tashin hankali. Ana nuna wannan a cikin Hoto na 3d, inda jirgin ƙasa mai shuɗi yana haifar da ƙarfin wutar lantarki na membrane don haɓakawa akan capacitor na membrane, yayin da koren bugun jini yana riƙe da ƙarfin ƙarfin membrane akai-akai.
Aiki na biyu mai mahimmanci na RRAM shine aiwatar da ma'aunin haɗi. Yin amfani da daidaitawar halayen analog na RRAM, haɗin I/O na iya yin nauyi daidai da haka. A gwaji na biyu, an tsara na'urar G0 zuwa matakai daban-daban na HCS, kuma an yi amfani da bugun bugun jini zuwa shigarwar Vin0. Ƙunƙarar shigar da bugun jini yana zana na yanzu (Iweight) daga na'urar, wanda yayi daidai da gudanarwa da madaidaicin yuwuwar faɗuwar Vtop - Vbot. Ana shigar da wannan ma'aunin nauyi na halin yanzu a cikin synapses na DPI da ƙwayoyin fitarwa na LIF. An yi rikodin ƙarfin lantarki na membrane na ƙwayoyin fitarwa ta amfani da oscilloscope kuma an nuna su a cikin siffa 3d. Ƙwaƙwalwar ƙarfin lantarki na membrane neuron don mayar da martani ga bugun bugun jini guda ɗaya ya yi daidai da tafiyar da ƙwaƙwalwar juriya, yana nuna cewa ana iya amfani da RRAM azaman nau'in nau'i na nau'in synaptic. Waɗannan gwaje-gwaje na farko guda biyu sun nuna cewa dandamalin neuromorphic na tushen RRAM yana iya aiwatar da abubuwan asali na ainihin hanyar Jeffress, wato layin jinkiri da da'ira mai gano daidaituwa. An gina dandalin kewayawa ta hanyar tara tubalan da ke jere gefe da gefe, kamar tubalan da ke cikin Hoto na 3b, da haɗa ƙofofinsu zuwa layin shigar gama gari. Mun tsara, ƙirƙira, da kuma gwada wani dandamali na neuromorphic wanda ya ƙunshi nau'i-nau'i masu fitarwa guda biyu suna karɓar bayanai guda biyu (Fig. 4a). Ana nuna zanen kewayawa a hoto na 4b. Matrix na 2 × 2 RRAM na sama yana ba da damar shigar da bugun jini zuwa ga jigon fitarwa guda biyu, yayin da ƙananan matrix 2 × 2 yana ba da damar haɗin kai na neurons biyu (N0, N1). Mun nuna cewa za'a iya amfani da wannan dandamali tare da daidaitawar layi na jinkiri da kuma ayyuka guda biyu na daidaitattun daidaituwa, kamar yadda aka nuna ta hanyar ma'auni na gwaji a cikin Fig. 4c-e.
Zane-zane da aka kirkira ta nau'ikan nau'ikan fitarwa guda biyu N0 da N1 suna karɓar bayanai guda biyu 0 da 1. Na'urori huɗu na sama na tsararrun suna bayyana haɗin haɗin gwiwar synaptic daga shigarwa zuwa fitarwa, kuma sel huɗu na ƙasa suna bayyana haɗin kai akai-akai tsakanin neurons. RRAMs masu launin suna wakiltar na'urorin da aka saita a cikin HCS a hannun dama: na'urorin da ke cikin HCS suna ba da damar haɗi kuma suna wakiltar ma'auni, yayin da na'urorin da ke cikin LCS suna toshe bugun bugun jini kuma suna kashe haɗin haɗi zuwa abubuwan fitarwa. b Zane na kewaye (a) tare da nau'ikan RRAM guda takwas waɗanda aka haskaka da shuɗi. c Ana yin layukan jinkiri ta hanyar amfani da kuzarin DPI synapses da LIF neurons. An saita RRAM kore don gudanar da aiki mai girma don samun damar haifar da glitch a fitarwa bayan jinkirin shigarwar Δt. d Misalin tsari na gano CD ɗin da ba shi da ma'ana na sigina masu dogaro da lokaci. Fitowar neuron 1, N1, yana kunna abubuwan shigar 0 da 1 tare da ɗan gajeren jinkiri. e Direction m CD circuit, da'irar da ke gano lokacin shigar da 1 ya kusanci shigarwar 0 kuma yana zuwa bayan shigarwar 0. Fitowar da'irar tana wakiltar neuron 1 (N1).
Layin jinkiri (Hoto 4c) kawai yana amfani da ƙarfin hali na DPI synapses da LIF neurons don sake haifar da karuwar shigarwa daga Vin1 zuwa Vout1 ta jinkirta Tdel. G3 RRAM da aka haɗa da Vin1 da Vout1 ne kawai aka tsara su a cikin HCS, sauran RRAMs an tsara su a cikin LCS. An tsara na'urar G3 don 92.6 µs don tabbatar da cewa kowane bugun jini na shigarwa yana ƙara ƙarfin lantarki na membran fitarwa sosai don isa bakin kofa kuma ya haifar da bugun bugun jini mai jinkiri. An ƙayyade jinkirin Tdel ta hanyar synaptik da madaidaitan lokacin jijiya. Masu gano daidaituwa suna gano abin da ya faru na alaƙa na ɗan lokaci amma siginar shigar da aka rarraba a sarari. CD ɗin da ba shi da ma'ana ya dogara ne akan abubuwan da aka shigar da su guda ɗaya waɗanda ke haɗuwa zuwa jijiya na fitarwa na gama gari (Hoto na 4d). RRAM biyun da ke haɗa Vin0 da Vin1 zuwa Vout1, G2 da G4 bi da bi an tsara su don babban gudanarwa. Zuwan spikes na lokaci guda akan Vin0 da Vin1 yana ƙara ƙarfin lantarki na membran neuron N1 sama da iyakar da ake buƙata don samar da karuwar fitarwa. Idan abubuwan shigar guda biyu sun yi nisa sosai cikin lokaci, cajin wutar lantarki da aka tara ta hanyar shigarwar farko na iya samun lokacin ruɓewa, yana hana yuwuwar membrane N1 isa ga ƙimar kofa. G1 da G2 an tsara su don kusan 65µs, wanda ke tabbatar da cewa haɓakar shigarwa guda ɗaya baya ƙara ƙarfin ƙarfin membrane isa ya haifar da haɓakar fitarwa. Gano daidaituwa tsakanin abubuwan da aka rarraba a sararin samaniya da lokaci shine ainihin aiki da ake amfani da shi a cikin ayyuka masu yawa na ji kamar guje wa cikas na tushen kwararar gani da gano tushen sauti. Don haka, faifan CD ɗin da ba su da alkibla da ƙima shine ginshiƙin ginin tushe don gina tsarin gani da sauti. Kamar yadda aka nuna ta halayen ma'auni na lokaci (duba ƙarin siffa 2), da'irar da aka tsara tana aiwatar da kewayon da ya dace na oda huɗu na ma'auni na lokaci. Don haka, yana iya lokaci guda ya dace da buƙatun tsarin gani da sauti. CD mai hankali na kai-tsaye, da'ira ce mai kula da tsarin sararin samaniya na isowar bugun bugun jini: daga dama zuwa hagu da kuma akasin haka. Yana da tushe mai tushe a cikin cibiyar gano motsi na asali na tsarin gani na Drosophila, wanda aka yi amfani da shi don ƙididdige kwatancen motsi da gano haɗuwa62. Don cimma nasarar CD, mai hankali dole ne a umurce su biyu ga neurons biyu daban-daban (N0, N1) kuma dole ne a kafa hanyar haɗi tsakanin su (Fig .e). Lokacin da aka karɓi shigarwar farko, NO yana amsawa ta ƙara ƙarfin wutar lantarki a saman membran sa sama da ƙimar kofa da aika ƙara. Wannan taron fitarwa, bi da bi, yana ƙone N1 godiya ga haɗin kai da aka haskaka a cikin kore. Idan taron shigarwa Vin1 ya zo ya ƙarfafa N1 yayin da ƙarfin lantarki na membrane yana da girma, N1 yana haifar da abin fitarwa wanda ke nuna cewa an sami wasa tsakanin abubuwan shigar guda biyu. Haɗin kai tsaye yana ba da damar N1 don fitar da fitarwa kawai idan shigarwar 1 ta zo bayan shigarwar 0. G0, G3, da G7 an tsara su zuwa 73.5 µS, 67.3 µS, da 40.2 µS, bi da bi, tabbatar da cewa ƙuri'a ɗaya akan shigarwar Vin0 yana haifar da jinkiri. fitowar fitarwa, yayin da N1's membrane yuwuwar kai ga ƙofa ne kawai lokacin da fashewar shigarwar biyu suka zo daidai. .
Bambance-bambance shine tushen ajizanci a cikin tsarin tsarin neuromorphic da aka tsara63,64,65. Wannan yana haifar da halayen nau'ikan neurons da synapses. Misalai na irin wannan rashin amfani sun haɗa da 30% (ma'anar daidaitaccen karkatacce) bambance-bambancen samun damar shigar da bayanai, daɗaɗɗen lokaci, da lokacin ɓata lokaci, don suna amma kaɗan (duba Hanyoyi). Wannan matsala ta fi fitowa fili idan aka haɗa nau'o'in jijiyoyi da yawa tare, kamar CD mai ma'ana wanda ya ƙunshi jijiya biyu. Don yin aiki yadda ya kamata, riba da lalata lokaci na jijiyoyin jijiyoyin biyu ya kamata su kasance iri ɗaya kamar yadda zai yiwu. Misali, babban bambance-bambancen samun shigar da bayanai na iya haifar da neuron guda ya wuce gona da iri ga bugun bugun jini yayin da sauran neuron ke da kyar. A kan fig. Hoto na 5a yana nuna cewa jijiya da aka zaɓa ba da gangan ba suna amsa daban-daban ga bugun bugun jini guda ɗaya. Wannan bambance-bambancen jijiyoyi yana da dacewa, misali, zuwa aikin CD masu hankali. A cikin tsarin da aka nuna a cikin fig. 5b, c, shigar da shigar da neuron 1 ya fi girma fiye da na neuron 0. Saboda haka, neuron 0 yana buƙatar nau'in shigarwa guda uku (maimakon 1) don isa bakin kofa, kuma neuron 1, kamar yadda aka sa ran, yana buƙatar abubuwan shigarwa guda biyu. Aiwatar da karu-lokaci mai dogaro da filastik biomimetic (STDP) wata hanya ce mai yuwuwar don rage tasirin rashin daidaituwa da sluggish jijiya da da'irori na synaptic akan aikin tsarin43. Anan muna ba da shawarar yin amfani da halayen filastik na ƙwaƙwalwar juzu'i azaman hanyar haifar da haɓaka haɓakar shigarwar jijiya da rage tasirin sauye-sauye a cikin da'irar neuromorphic. Kamar yadda aka nuna a cikin fig. 4e, matakan gudanarwa masu alaƙa da RRAM synaptic taro yadda ya kamata ya daidaita daidaitaccen ƙarfin ƙarfin lantarki na jijiyoyi. Muna amfani da dabarar shirye-shiryen RRAM mai jujjuyawa. Don shigarwar da aka bayar, ana sake tsara ƙimar gudanarwa na ma'aunin synaptic har sai an sami halayen da'irar (duba Hanyoyi).
Ma'auni na gwaji na martanin jijiya guda tara da aka zaɓa ba da gangan ba zuwa bugun bugun jini iri ɗaya. Amsar ta bambanta a cikin yawan jama'a, yana shafar ribar shigarwa da tsayayyen lokaci. b Ma'auni na gwaji na tasirin neurons akan sauye-sauye na neurons da ke shafar CD mai hankali. Jibiyoyin fitarwa na CD guda biyu masu hankali suna amsawa daban-daban don shigar da kuzari saboda canjin neuron-to-neuron. Neuron 0 yana da ƙarancin shigar da ƙara fiye da neuron 1, don haka yana ɗaukar nau'ikan shigarwa uku (maimakon 1) don ƙirƙirar ƙaruwar fitarwa. Kamar yadda aka zata, neuron 1 ya kai ƙofa tare da abubuwan shigarwa guda biyu. Idan shigarwar 1 ta isa Δt = 50 µs bayan gobarar neuron 0, CD ya yi shiru saboda Δt ya fi tsayin lokaci na neuron 1 (kimanin 22 µs). c yana raguwa da Δt = 20 µs, don haka shigar da 1 kololuwa lokacin da firing neuron 1 ya kasance har yanzu yana da girma, yana haifar da gano abubuwan shigarwa guda biyu lokaci guda.
Abubuwa biyu da aka yi amfani da su a cikin ginshiƙi na lissafin ITD sune layin jinkiri da CD ɗin da ba ya jin daɗi. Duk da'irori biyu suna buƙatar madaidaicin daidaitawa don tabbatar da kyakkyawan aikin sanya abu. Layin jinkiri dole ne ya isar da daidaitaccen sigar mafi girman shigarwar (Hoto 6a), kuma CD ɗin dole ne a kunna shi kawai lokacin da shigarwar ta faɗi cikin kewayon ganowa. Don layin jinkiri, ma'aunin synaptic na haɗin shigarwa (G3 a cikin siffa 4a) an sake tsara shi har sai an sami jinkirin manufa. Saita juriya a kusa da jinkirin da aka yi niyya don dakatar da shirin: ƙarami juriya, mafi wahala shine samun nasarar saita layin jinkiri. A kan fig. Hoto na 6b yana nuna sakamakon tsarin daidaita layin jinkirta: ana iya ganin cewa shirin da aka tsara zai iya samar da duk jinkirin da ake buƙata a cikin tsarin ƙira (daga 10 zuwa 300 μs). Matsakaicin adadin maimaitawa na daidaitawa yana rinjayar ingancin tsarin daidaitawa: 200 na iya rage kuskuren zuwa ƙasa da 5%. Ƙimar daidaitawa ɗaya ta yi daidai da saiti/sake saitin aiki na tantanin halitta RRAM. Hakanan tsarin daidaitawa yana da mahimmanci don haɓaka daidaiton ƙirar CD ɗin gano abin da ya faru nan take. An ɗauki maimaita juzu'i goma don cimma ingantacciyar ƙima na gaske (watau ƙimar abubuwan da aka gano daidai da dacewa) sama da 95% (layin shuɗi a cikin hoto na 6c). Duk da haka, tsarin daidaitawa bai shafi abubuwan da suka faru na gaskiya ba (wato, yawan abubuwan da suka faru da kuskuren da suka dace). Wata hanyar da aka lura a cikin tsarin ilimin halitta don shawo kan ƙayyadaddun lokaci na hanyoyi masu saurin kunnawa shine sakewa (wato, yawancin kwafi na abu ɗaya ana amfani dashi don yin aikin da aka ba). Kwarewar ilimin halitta66, mun sanya da'irori na CD da yawa a cikin kowane nau'in CD tsakanin layin jinkiri guda biyu don rage tasirin tabbataccen ƙarya. Kamar yadda aka nuna a cikin fig. 6c (layin kore), sanya abubuwan CD guda uku a cikin kowane tsarin CD na iya rage ƙimar ƙararrawar ƙarya zuwa ƙasa da 10-2.
Tasirin sauye-sauyen neuronal akan da'irar layin jinkiri. b Za a iya daidaita da'irar layin jinkiri zuwa manyan jinkiri ta hanyar saita madaidaitan lokutan jigon LIF masu dacewa da synapses na DPI zuwa manyan dabi'u. Ƙara yawan maimaitawa na tsarin daidaitawa na RRAM ya sa ya yiwu a inganta daidaito na jinkirin da aka yi niyya: 200 maimaitawa ya rage kuskuren zuwa kasa da 5%. Juyawa ɗaya yayi daidai da aikin SET/SAKESET akan tantanin halitta RRAM. Ana iya aiwatar da kowane nau'in CD a cikin samfurin c Jeffress ta amfani da abubuwan CD masu kamanceceniya da N don ƙarin sassauci dangane da gazawar tsarin. d Ƙarin gyare-gyare na RRAM yana haɓaka ƙimar gaskiya ta gaskiya (layin shuɗi), yayin da ƙimar ƙimar ƙarya ta kasance mai zaman kanta daga adadin iterations (layin kore). Sanya ƙarin abubuwan CD a layi daya yana guje wa gano kuskuren matches na CD.
Yanzu muna kimanta aiki da amfani da wutar lantarki na tsarin haɗaɗɗen abu na ƙarshe zuwa ƙarshe wanda aka nuna a cikin Hoto 2 ta amfani da ma'auni na abubuwan faɗakarwa na firikwensin pMUT, CD, da da'irar layin jinkiri waɗanda ke haɗa jadawali na lissafin neuromorphic. Misalin Jeffress (Fig. 1a). Dangane da jadawali na lissafin neuromorphic, mafi girman adadin CD ɗin, mafi kyawun ƙudurin kusurwa, amma kuma mafi girman ƙarfin tsarin (Fig. 7a). Ana iya samun sulhu ta hanyar kwatanta daidaiton daidaikun abubuwan da aka haɗa (pMUT firikwensin, neurons, da da'irori na synaptic) tare da daidaiton tsarin gaba ɗaya. Ƙaddamarwar layin jinkiri yana iyakancewa ta lokacin da aka kwatanta da synapses da neurons, wanda a cikin tsarinmu ya wuce 10 µs, wanda ya dace da ƙudurin kusurwa na 4 ° (duba Hanyoyi). Ƙarin ci gaba na nodes tare da fasahar CMOS za su ba da izinin ƙirar jijiyoyi da na'urorin synaptic tare da ƙananan lokutan lokaci, yana haifar da daidaito mafi girma na abubuwan layin jinkiri. Duk da haka, a cikin tsarin mu, daidaito yana iyakance ta kuskuren pMUT a kimanta matsayi na kusurwa, watau 10 ° (layin kwance mai launin shuɗi a cikin Fig. 7a). Mun gyara adadin na'urorin CD a 40, wanda yayi daidai da ƙudurin kusurwa na kusan 4 °, watau daidaitaccen angular na jadawali (layin kwance mai haske blue a cikin siffa 7a). A matakin tsarin, wannan yana ba da ƙuduri na 4 ° da daidaito na 10 ° don abubuwan da ke 50 cm a gaban tsarin firikwensin. Wannan kimar tana kwatankwacinta da tsarin gurɓataccen sauti na neuromorphic da aka ruwaito a ref. 67. Ana iya samun kwatancen tsarin da aka tsara tare da yanayin fasaha a cikin Ƙarin Teburin 1. Ƙara ƙarin pMUTs, haɓaka matakin siginar sauti, da rage sautin lantarki shine hanyoyin da za a iya kara inganta daidaitattun wuri. ) an kiyasta a 9.7. nz. 55. An ba da raka'o'in CD 40 akan jadawali na lissafi, simintin SPICE ya ƙididdige ƙarfin kowane aiki (watau makamashi sakawa abu) ya zama 21.6 nJ. Ana kunna tsarin neuromorphic ne kawai lokacin da abin shigarwa ya zo, watau lokacin da igiyar murya ta isa kowane mai karɓar pMUT kuma ya wuce matakin ganowa, in ba haka ba ya kasance mara aiki. Wannan yana guje wa amfani da wutar da ba dole ba lokacin da babu siginar shigarwa. Yin la'akari da mitar ayyukan gida na 100 Hz da lokacin kunnawa na 300 µs a kowane aiki (matsakaicin yuwuwar ITD), ikon amfani da jadawali na lissafin neuromorphic shine 61.7 nW. Tare da pre-processing neuromorphic da aka yi amfani da shi ga kowane mai karɓar pMUT, yawan wutar lantarki na tsarin duka ya kai 81.6 nW. Don fahimtar ingancin makamashi na tsarin neuromorphic da aka tsara idan aka kwatanta da kayan aiki na al'ada, mun kwatanta wannan lambar zuwa makamashin da ake buƙata don yin aiki iri ɗaya a kan ƙananan ƙarfin lantarki na zamani ta amfani da ko dai neuromorphic ko na al'ada beamforming68 Skill. Hanyar neuromorphic tana la'akari da mataki na analog-to-dijital Converter (ADC), sannan sai kuma matatar band-pass da matakin cire ambulaf (hanyar Teeger-Kaiser). A ƙarshe, ana yin aikin kofa don cire ToF. Mun tsallake lissafin ITD dangane da ToF da jujjuyawar zuwa kimanta matsayi na kusurwa tunda wannan yana faruwa sau ɗaya don kowane ma'auni (duba Hanyoyi). Yin la'akari da ƙimar ƙima na 250 kHz akan tashoshi biyu (masu karɓar pMUT), 18 band wucewa ayyukan tacewa, ayyukan cire ambulaf 3, da aiki na ƙofa 1 a kowane samfurin, ana ƙididdige yawan amfani da wutar lantarki a 245 microwatts. Wannan yana amfani da yanayin ƙarancin iko na microcontroller69, wanda ke kunna lokacin da algorithms ba sa aiwatarwa, wanda ke rage yawan wutar lantarki zuwa 10.8 µW. Amfanin wutar lantarki na maganin sarrafa siginar katako wanda aka gabatar a cikin tunani. 31, tare da masu karɓa na 5 pMUT da 11 katako da aka rarraba daidai a cikin jirgin azimuth [-50 °, + 50 °], shine 11.71 mW (duba sashin Hanyoyi don cikakkun bayanai). Bugu da kari, muna ba da rahoton amfani da wutar lantarki na tushen FPGA47 na tushen Lokaci Difference Encoder (TDE) wanda aka kiyasta a 1.5 mW a matsayin maye gurbin samfurin Jeffress don gano abu. Dangane da waɗannan ƙididdiga, tsarin tsarin neuromorphic da aka tsara yana rage yawan amfani da wutar lantarki ta umarni biyar na girma idan aka kwatanta da na'ura mai sarrafawa ta hanyar amfani da dabarun ƙirar katako na gargajiya don ayyukan gano abubuwa. Ɗauki hanyar neuromorphic don sarrafa sigina akan na'urar sarrafa microcontroller na yau da kullun yana rage yawan amfani da wutar lantarki da kusan umarni biyu na girma. Ana iya bayanin tasirin tsarin da aka tsara ta hanyar haɗin da'irar asynchronous resistive-memory analog da'irar mai iya yin ƙididdige ƙididdiga a cikin ƙwaƙwalwar ajiya da kuma rashin canjin analog-zuwa-dijital da ake buƙata don gane sigina.
Ƙaddamar da Angular (blue) da kuma amfani da wutar lantarki (kore) na aikin ƙaddamarwa ya dogara da adadin nau'in CD. Barin kwance mai duhu shuɗi yana wakiltar daidaiton kusurwar PMUT kuma shuɗin kwance mai haske yana wakiltar daidaiton kusurwar jadawali na lissafin neuromorphic. b Yin amfani da wutar lantarki na tsarin da aka tsara da kwatanta tare da aiwatar da microcontroller guda biyu da aka tattauna da aiwatar da dijital na Difference Encoder (TDE) 47 FPGA.
Don rage yawan amfani da wutar lantarki na tsarin ƙaddamar da manufa, mun ɗauki ciki, tsarawa da aiwatar da ingantaccen, da'irar neuromorphic na tushen RRAM wanda ke aiwatar da bayanan siginar da na'urori masu auna firikwensin suka haifar don ƙididdige matsayin abin da ake nufi a zahiri. lokaci. . Duk da yake hanyoyin sarrafa al'ada suna ci gaba da yin samfurin siginar da aka gano da yin ƙididdiga don fitar da bayanai masu amfani, ƙirar neuromorphic da aka gabatar tana aiwatar da lissafin asynchronously yayin da bayanai masu amfani suka zo, suna haɓaka ƙarfin tsarin ta hanyar umarni biyar na girma. Bugu da ƙari, muna haskaka sassauƙar da'irar neuromorphic tushen RRAM. Ƙarfin RRAM don canza ɗawainiya ta hanyar da ba ta da ƙarfi (plasticity) tana ramawa ga bambance-bambancen da'irar wutar lantarki mai ƙarancin ƙarfi na DPI na synaptic da na jijiyoyi. Wannan ya sa wannan da'irar tushen RRAM ta zama mai ƙarfi da ƙarfi. Manufar mu ba shine mu cire hadaddun ayyuka ko alamu daga sigina ba, amma don gano abubuwa a ainihin lokacin. Tsarin mu kuma zai iya danne siginar da inganci kuma a ƙarshe aika shi zuwa ƙarin matakan sarrafawa don yanke shawara mai rikitarwa lokacin da ake buƙata. A cikin mahallin aikace-aikacen yanki, matakin ƙaddamar da aikin mu na neuromorphic zai iya ba da bayani game da wurin da abubuwa suke. Ana iya amfani da wannan bayanin, alal misali, don gano motsi ko ganewar motsi. Muna jaddada mahimmancin haɗa na'urori masu ƙarancin ƙarfi kamar pMUT tare da ƙananan wutar lantarki. Don wannan, hanyoyin neuromorphic sun kasance maɓalli yayin da suka jagoranci mu don haɓaka sabbin ayyukan da'ira na hanyoyin ƙididdigewa na ilimin halitta kamar samfurin Jeffress. A cikin mahallin aikace-aikacen haɗakar firikwensin, ana iya haɗa tsarin mu tare da na'urori masu auna firikwensin taron daban-daban don samun ƙarin ingantattun bayanai. Ko da yake mujiya suna da kyau wajen gano ganima a cikin duhu, suna da kyakkyawan gani kuma suna yin binciken ji da gani da ido kafin su kama ganima70. Lokacin da wani nau'in naurar jijiya na musamman ya kunna wuta, mujiya tana karɓar bayanan da yake buƙata don tantance ta wacce hanya za ta fara binciken na gani, don haka yana mai da hankali kan ɗan ƙaramin abin da ake gani. Haɗin na'urori masu auna firikwensin gani (kamara DVS) da na'urar firikwensin sauraro (bisa pMUT) yakamata a bincika don haɓaka wakilai masu cin gashin kansu na gaba.
Na'urar firikwensin pMUT yana kan PCB tare da masu karɓa guda biyu kusan 10 cm tsakanin su, kuma mai watsawa yana tsakanin masu karɓa. A cikin wannan aikin, kowane membrane shine tsarin bimorph da aka dakatar wanda ya ƙunshi nau'i biyu na piezoelectric aluminum nitride (AlN) 800 nm lokacin farin ciki sandwiched tsakanin nau'i uku na molybdenum (Mo) 200 nm lokacin farin ciki kuma an rufe shi da Layer 200 nm lokacin farin ciki. saman passivating SiN Layer kamar yadda aka bayyana a cikin tunani. 71. Ana amfani da na'urori na ciki da na waje zuwa kasa da saman saman molybdenum, yayin da na'urar lantarki na molybdenum na tsakiya ba shi da tsari kuma ana amfani da shi azaman ƙasa, wanda ya haifar da membrane mai nau'i-nau'i guda hudu.
Wannan gine-ginen yana ba da damar yin amfani da nakasar membrane na kowa, yana haifar da ingantaccen watsawa da karɓar hankali. Irin wannan pMUT yawanci yana nuna motsin hankali na 700 nm/V azaman emitter, yana samar da matsa lamba na 270 Pa/V. A matsayin mai karɓa, fim ɗin pMUT ɗaya yana nuna ɗan gajeren kewayawa na 15 nA/Pa, wanda ke da alaƙa kai tsaye da haɗin gwiwar piezoelectric na AlN. Canjin fasaha na ƙarfin lantarki a cikin AlN Layer yana haifar da canji a cikin mitar resonant, wanda za'a iya biya ta hanyar yin amfani da rashin daidaituwa na DC zuwa pMUT. An auna ƙarfin DC a 0.5 kHz/V. Don halayyar sauti, ana amfani da makirufo a gaban pMUT.
Don auna bugun bugun echo, mun sanya farantin rectangular tare da yanki na kusan 50 cm2 a gaban pMUT don yin nuni da raƙuman sauti da aka fitar. Duka nisa tsakanin faranti da kusurwar dangi da jirgin pMUT ana sarrafa su ta amfani da masu riƙewa na musamman. Tushen wutar lantarki na Tectronix CPX400DP yana nuna son kai ga membranes pMUT uku, yana daidaita mitar resonant zuwa 111.9 kHz31, yayin da masu watsawa ke motsa su ta hanyar janareta mai bugun bugun jini na Tectronix AFG 3102 wanda aka kunna zuwa mitar resonant (111.9 kHz) da zagayowar aiki na 0.0. Matsalolin da ake karantawa daga tashoshin fitarwa guda huɗu na kowane mai karɓar pMUT ana canza su zuwa ƙarfin lantarki ta amfani da na'ura mai ban sha'awa na musamman da na lantarki, kuma siginar da aka samu ana ƙididdige su ta hanyar tsarin sayan bayanan Spektrum. Iyakar ganowa an kwatanta shi ta hanyar siginar pMUT a ƙarƙashin yanayi daban-daban: mun matsar da mai nunawa zuwa nisa daban-daban [30, 40, 50, 60, 80, 100] cm kuma mun canza kusurwar tallafin pMUT ([0, 20, 40] o). ) Hoto 2b yana nuna ƙudurin gano ITD na ɗan lokaci dangane da daidaitaccen matsayi na kusurwa a cikin digiri.
Wannan labarin yana amfani da da'irori na RRAM daban-daban guda biyu. Na farko jeri ne na na'urori 16,384 (16,000) (na'urori 128 × 128) a cikin tsarin 1T1R tare da transistor daya da resistor daya. Guntu na biyu shine dandalin neuromorphic da aka nuna a cikin siffa 4a. Tantanin halitta na RRAM ya ƙunshi fim ɗin HfO2 mai kauri mai girman 5 nm wanda aka saka a cikin tarin TiN/HfO2/Ti/TiN. An haɗa tarin RRAM a cikin layin baya (BEOL) na daidaitaccen tsari na 130nm CMOS. Da'irori neuromorphic na tushen RRAM suna gabatar da ƙalubalen ƙira don duk tsarin lantarki na analog wanda na'urorin RRAM ke rayuwa tare da fasahar CMOS ta gargajiya. Musamman, yanayin tafiyar da na'urar RRAM dole ne a karanta kuma a yi amfani da shi azaman canjin aiki na tsarin. Don wannan, an ƙirƙira da'ira, ƙirƙira da gwadawa wanda ke karanta halin yanzu daga na'urar lokacin da aka karɓi bugun bugun jini kuma yana amfani da wannan halin yanzu don auna martanin synapse na bambance-bambancen mahaɗan biyu (DPI). Ana nuna wannan da'ira a cikin Hoto na 3a, wanda ke wakiltar ainihin tubalan ginin dandali na neuromorphic a cikin Hoto 4a. Buga bugun jini yana kunna ƙofar na'urar 1T1R, yana haifar da halin yanzu ta hanyar RRAM daidai da sarrafa na'urar G (Iweight = G (Vtop - Vx)). Shigar da jujjuyawar da'irar amplifier mai aiki (op-amp) tana da madaurin wutar lantarki na Vtop na DC akai-akai. Ra'ayin mara kyau na op-amp zai samar da Vx = Vtop ta samar da daidaitaccen halin yanzu daga M1. Iweight na yanzu da aka samo daga na'urar ana allura a cikin synapse na DPI. Ƙarfin halin yanzu zai haifar da ƙarin lalacewa, don haka gudanar da RRAM yana aiwatar da ma'aunin synaptic yadda ya kamata. Wannan madaidaicin halin yanzu ana yin allurar ta hanyar membrane capacitor na Leaky Integration and Excitation (LIF) neurons, inda aka haɗa shi azaman ƙarfin lantarki. Idan an shawo kan madaidaicin wutar lantarki na membrane (ƙarashin wutar lantarki na inverter), ana kunna ɓangaren fitarwa na neuron, yana samar da haɓakar fitarwa. Wannan bugun jini yana dawowa kuma yana jujjuya capacitor membrane na neuron zuwa ƙasa, yana haifar da fitarwa. Ana ƙara wannan kewayawa tare da mai faɗaɗa bugun jini (ba a nuna shi a cikin siffa 3a ba), wanda ke siffata bugun bugun jini na LIF neuron zuwa nisan bugun bugun jini. Ana kuma gina na'urori masu yawa a cikin kowane layi, suna ba da damar yin amfani da wutar lantarki zuwa sama da ƙasa na na'urar RRAM.
Gwajin wutar lantarki ya haɗa da yin nazari da yin rikodin halayen da'irar analog, da kuma shirye-shirye da karanta na'urorin RRAM. Duk matakan biyu suna buƙatar kayan aiki na musamman, waɗanda duk an haɗa su da allon firikwensin a lokaci guda. Samun damar zuwa na'urorin RRAM a cikin da'irori neuromorphic ana aiwatar da su daga kayan aikin waje ta hanyar multixer (MUX). MUX yana raba tantanin halitta 1T1R daga sauran kewayen da yake nasa, yana ba da damar karantawa da/ko tsara na'urar. Don tsarawa da karanta na'urorin RRAM, ana amfani da na'urar Keithley 4200 SCS tare da haɗin gwiwar Arduino microcontroller: na farko don ingantaccen tsarar bugun jini da karatun yanzu, kuma na biyu don saurin samun dama ga abubuwan 1T1R guda ɗaya a cikin tsarin ƙwaƙwalwar ajiya. Aikin farko shine samar da na'urar RRAM. Ana zaɓar ƙwayoyin sel ɗaya bayan ɗaya kuma ana amfani da ingantaccen ƙarfin lantarki tsakanin sama da ƙasa. A wannan yanayin, halin yanzu yana iyakance ga jerin dubun microamperes saboda samar da wutar lantarki mai dacewa ga mai zaɓin transistor. Tantanin halitta na RRAM zai iya zagayawa tsakanin ƙaramin hali (LCS) da babban yanayin gudanarwa (HCS) ta amfani da RESET da SET ayyuka, bi da bi. Ana aiwatar da aikin da aka saita ta hanyar amfani da bugun murabba'i mai kusurwa tare da tsawon lokaci na 1 μs da kuma ƙarfin ƙarfin lantarki, kuma sync bugun zuciya na 0.9-1.3 v to kofar zabar transistor. Waɗannan dabi'un suna ba da damar daidaita ayyukan RRAM a tazara 20-150µs. Don SAKESET, 1µs mai faɗi, 3V mafi girma bugun bugun jini ana amfani da shi zuwa ƙasan lantarki (layin bit) na tantanin halitta lokacin da ƙarfin ƙofar yana cikin kewayon 2.5-3.0 V. Abubuwan shigarwa da abubuwan da ake fitarwa na da'irorin analog sune sigina masu ƙarfi. . Don shigar da bayanai, mun haɗu da na'urorin bugun bugun jini guda biyu na HP 8110 tare da masu samar da siginar Tektronix AFG3011. Ƙunshin shigar da bugun jini yana da faɗin 1µs da tsayi/faɗi gefen 50 ns. Irin wannan nau'in bugun jini ana ɗauka ya zama ƙulli a cikin da'irar tushen glitch na analog. Dangane da siginar fitarwa, an yi rikodin siginar fitarwa ta amfani da Teledyne LeCroy 1 GHz oscilloscope. An tabbatar da saurin sayan oscilloscope ba zai zama abin iyakancewa ba a cikin bincike da kuma samun bayanan da'ira.
Yin amfani da kuzarin lantarki na analog don yin kwaikwayon halayen neurons da synapses shine kyakkyawan tsari kuma ingantaccen bayani don haɓaka haɓakar lissafi. Babban hasara na wannan ƙirar ƙididdiga shine cewa zai bambanta daga tsari zuwa tsari. Mun ƙididdige bambance-bambancen neurons da da'irori na synaptic (Ƙarin Hoton 2a,b). Daga cikin dukkanin bayyanar cututtuka na sauye-sauye, waɗanda ke da alaƙa da ma'auni na lokaci da kuma shigar da shigarwa suna da tasiri mafi girma a matakin tsarin. Tsawon lokaci na LIF neuron da DPI synapse an ƙaddara ta hanyar RC, inda darajar R ke sarrafawa ta hanyar ƙarfin lantarki da aka yi amfani da shi a ƙofar transistor (Vlk don neuron da Vtau don synapse), ƙayyade yawan yabo. An ayyana samun shigar da ƙara azaman ƙarfin ƙarfin ƙarfin da na'urorin synaptic da neuronal membrane capacitors suka ƙarfafa ta bugun bugun bugun jini. Ana sarrafa ribar shigarwa ta wani transistor na son rai wanda ke daidaita abubuwan shigar da halin yanzu. An yi wani simintin Monte Carlo wanda aka daidaita akan tsarin ST Microelectronics' 130nm don tattara wasu ribar shigar da ƙididdiga ta dindindin. Ana gabatar da sakamakon a Ƙarin Hoto na 2, inda aka ƙididdige ribar shigarwar da madaidaicin lokaci a matsayin aikin ƙarfin wutar lantarki na son rai da ke sarrafa ƙimar ɗigo. Alamun kore suna ƙididdige madaidaicin karkatar da lokaci daga ma'ana. Dukansu neurons da da'irori na synaptic sun sami damar bayyana kewayon madaidaitan lokutan lokaci a cikin kewayon 10-5-10-2 s, kamar yadda aka nuna a Ƙarin Tsarin Hoto. Ƙaddamar da shigarwa (Ƙarin Hoton 2e, d) na sauye-sauyen neuronal da synapse ya kasance kusan 8% da 3%, bi da bi. Irin wannan rashi yana da kyau a rubuce a cikin wallafe-wallafe: an yi ma'auni daban-daban akan tsararrun kwakwalwan DYNAP don tantance rashin daidaituwa tsakanin yawan LIF63 neurons. An auna synapses a cikin guntun siginar gauraya na BrainScale kuma an bincika rashin daidaituwarsu, kuma an ba da shawarar hanyar daidaitawa don rage tasirin canjin matakin-tsari64.
Ayyukan RRAM a cikin da'irori na neuromorphic sau biyu ne: ma'anar gine-gine (ma'anar hanyar sadarwa zuwa kayan aiki) da aiwatar da ma'aunin synaptic. Ana iya amfani da kadarorin na ƙarshe don magance matsalar sauye-sauyen da'irori na neuromorphic. Mun ƙirƙiri hanyar daidaitawa mai sauƙi wanda ya haɗa da sake tsara na'urar RRAM har sai an bincika da'irar ta cika wasu buƙatu. Don shigarwar da aka bayar, ana lura da fitarwa kuma ana sake tsara RRAM har sai an cimma halayen da aka yi niyya. An gabatar da lokacin jira na s 5 tsakanin ayyukan shirye-shirye don magance matsalar shakatawa na RRAM wanda ke haifar da jujjuyawar tafiyar lokaci (Ƙarin Bayani). Ana daidaita ma'aunin synaptic ko daidaita daidai da buƙatun da'irar neuromorphic da ake ƙira. An taƙaita tsarin daidaitawa a cikin ƙarin algorithms [1, 2] waɗanda ke mai da hankali kan mahimman abubuwa guda biyu na dandamali na neuromorphic, layin jinkiri da CD mara hankali. Don kewayawa tare da layin jinkiri, halayen manufa shine samar da bugun bugun jini tare da jinkirin Δt. Idan ainihin jinkirin da'ira ya yi ƙasa da ƙimar manufa, yakamata a rage nauyin synaptic na G3 (ya kamata a sake saita G3 sannan a saita zuwa ƙaramin madaidaicin Icc na yanzu). Sabanin haka, idan ainihin jinkirin ya fi ƙimar manufa, dole ne a ƙara yawan tafiyar da G3 (dole ne a sake saita G3 sannan a saita zuwa ƙimar Icc mafi girma). Ana maimaita wannan tsari har sai jinkirin da kewaye ya haifar ya dace da ƙimar manufa kuma an saita haƙuri don dakatar da tsarin daidaitawa. Don CD ɗin da ba su da hankali, na'urorin RRAM guda biyu, G1 da G3, suna cikin tsarin daidaitawa. Wannan kewayawa yana da abubuwan shigar guda biyu, Vin0 da Vin1, jinkiri ta dt. Kewayar ya kamata kawai ta amsa jinkirin da ke ƙasa da kewayon daidaitawa [0,dtCD]. Idan babu kololuwar fitarwa, amma kololuwar shigar yana kusa, ya kamata a haɓaka na'urorin RRAM biyu don taimakawa neuron ya isa bakin kofa. Sabanin haka, idan da'irar ta amsa jinkirin da ya wuce iyakar dtCD, dole ne a rage yawan tafiyarwa. Maimaita tsarin har sai an sami daidaitaccen hali. Ana iya daidaita halin yanzu yarda ta hanyar ginanniyar da'irar analog a cikin ref. 72.73. Tare da wannan da'irar da aka gina, ana iya aiwatar da irin waɗannan hanyoyin lokaci-lokaci don daidaita tsarin ko sake amfani da shi don wani aikace-aikacen.
Muna ƙididdige yawan wutar lantarki na tsarin sarrafa siginar mu na neuromorphic akan daidaitaccen 32-bit microcontroller68. A cikin wannan kimantawa, muna ɗaukar aiki tare da saiti ɗaya kamar a cikin wannan takarda, tare da mai watsa pMUT ɗaya da masu karɓar pMUT guda biyu. Wannan hanyar tana amfani da matatar bandpass, sannan kuma matakin cire ambulaf (Teeger-Kaiser), sannan kuma a ƙarshe ana amfani da aikin ƙofa ga siginar don cire lokacin tashi. An bar lissafin ITD da jujjuyawar sa zuwa kusurwoyin ganowa a cikin kimantawa. Muna yin la'akari da aiwatar da matattarar bandeji ta amfani da tsari na 4 mara iyaka mara iyaka ta amsa matatar da ke buƙatar ayyuka masu iyo 18. Cire ambulaf yana amfani da ƙarin ayyuka uku masu iyo, kuma ana amfani da aiki na ƙarshe don saita ƙofa. Ana buƙatar jimlar ayyuka 22 masu iyo don aiwatar da siginar. Siginar da aka watsa shine ɗan gajeren fashe na 111.9 kHz sine waveform wanda aka samar kowane ms 10 wanda ya haifar da mitar aiki na 100 Hz. Mun yi amfani da ƙimar samfurin 250 kHz don biyan Nyquist da taga 6 ms don kowane ma'auni don ɗaukar kewayon mita 1. Lura cewa miliyon 6 shine lokacin tashin wani abu mai nisan mita 1. Wannan yana ba da amfani da wutar lantarki na 180 µW don juyawa A/D a 0.5 MSPS. Gabatarwar siginar ita ce 6.60 MIPS (umarni a sakan daya), yana samar da 0.75mW. Koyaya, microcontroller na iya canzawa zuwa yanayin ƙarancin wutar lantarki 69 lokacin da algorithm baya gudana. Wannan yanayin yana ba da amfani mai ƙarfi na 10.8 μW da lokacin farkawa na 113 μs. Idan aka ba da mitar agogo na 84 MHz, microcontroller yana kammala duk ayyukan neuromorphic algorithm a cikin 10 ms, kuma algorithm yana ƙididdige zagayowar aiki na 6.3%, don haka yana amfani da yanayin ƙarancin ƙarfi. Sakamakon wutar lantarki shine 244.7 μW. Lura cewa mun watsar da fitowar ITD daga ToF da jujjuyawar zuwa kusurwar ganowa, don haka rage ƙimar amfani da wutar lantarki na microcontroller. Wannan yana ba da ƙarin ƙima don ingantaccen makamashi na tsarin da aka tsara. A matsayin ƙarin yanayin kwatanta, muna ƙididdige amfani da wutar lantarki na hanyoyin ƙirar katako na gargajiya da aka gabatar a cikin tunani. 31.54 lokacin da aka saka shi a cikin wannan microcontroller68 a wutar lantarki ta 1.8V. Ana amfani da membran pMUT guda biyar daidai gwargwado don samun bayanai don yin katako. Dangane da sarrafa kanta, hanyar samar da katako da ake amfani da ita ita ce taƙaitawar jinkiri. Ya ƙunshi kawai yin amfani da jinkiri zuwa hanyoyin da ya dace da bambancin da ake tsammani a lokutan isowa tsakanin layi ɗaya da hanyar tunani. Idan siginonin suna cikin lokaci, jimlar waɗannan sigina za su sami ƙarfi mai ƙarfi bayan canjin lokaci. Idan sun fita daga lokaci, tsangwama mai lalacewa zai iyakance ƙarfin jimlar su. cikin dangantaka. A kan fig. 31, an zaɓi ƙimar ƙima na 2 MHz don canza lokacin bayanai ta adadin samfuran samfuran. Ingantacciyar hanya mafi sauƙi ita ce kiyaye ƙimar samfurin 250 kHz da amfani da Maɓalli Mai Ƙarfi (FIR) tace don haɗa jinkirin juzu'i. Za mu ɗauka cewa rikitaccen tsarin ƙirar ƙirar ƙirar an ƙaddara shi ne ta hanyar canjin lokaci, tunda kowane tashoshi yana haɗuwa da matatar FIR tare da taps 16 a kowace hanya. Don ƙididdige adadin MIPS da ake buƙata don wannan aiki, muna yin la'akari da taga na 6ms kowace ma'auni don ɗaukar kewayon mita 1, tashoshi 5, kwatancen beamforming 11 (kewayon +/- 50 ° a cikin matakan 10 °). Ma'auni 75 a cikin daƙiƙa guda sun tura microcontroller zuwa iyakar MIPS 100. mahada. 68, wanda ya haifar da raguwar wutar lantarki na 11.26 mW don jimlar wutar lantarki na 11.71 mW bayan ƙara gudunmawar ADC a kan jirgin.
Ana samun bayanan da ke goyan bayan sakamakon wannan binciken daga marubucin, FM, bisa ga buƙatu mai ma'ana.
Indiveri, G. & Sandamirskaya, Y. Muhimmancin sararin samaniya da lokaci don sarrafa sigina a cikin ma'aikatan neuromorphic: ƙalubalen haɓaka ƙananan ƙarfi, masu cin gashin kansu waɗanda ke hulɗa da yanayi. Indiveri, G. & Sandamirskaya, Y. Muhimmancin sararin samaniya da lokaci don sarrafa sigina a cikin ma'aikatan neuromorphic: ƙalubalen haɓaka ƙananan ƙarfi, masu cin gashin kansu waɗanda ke hulɗa da yanayi.Indiveri G. da Sandamirskaya Y. Muhimmancin sararin samaniya da lokaci don sarrafa siginar siginar a cikin ma'aikatan neuromorphic: ƙalubalen haɓaka ƙananan masu cin gashin kansu da ke hulɗa da yanayi. Indiveri, G. & Sandamirskaya, Y.代理的挑战。 Indiveri, G. & Sandamirskaya, Y.Indiveri G. da Sandamirskaya Y. Muhimmancin sararin samaniya da lokaci don sarrafa siginar siginar a cikin ma'aikatan neuromorphic: ƙalubalen haɓaka ƙananan masu cin gashin kansu da ke hulɗa da yanayi.Gudanar da Siginar IEEE. Jaridar 36, 16-28 (2019).
Thorpe, Lokacin Zuwan Kololuwar SJ: Ingantacciyar Tsarin Lambar Hanyar Sadarwar Jijiya. a cikin Eckmiller, R., Hartmann, G. & Hauske, G. (eds). a cikin Eckmiller, R., Hartmann, G. & Hauske, G. (eds).a cikin Eckmiller, R., Hartmann, G. da Hauske, G. (eds.).A cikin Eckmiller, R., Hartmann, G., da Hauske, G. (eds.). Daidaita aiki a cikin tsarin jijiya da kwamfutoci 91-94 (North-Holland Elsevier, 1990).
Levy, WB & Calvert, VG Sadarwa yana cinye makamashi sau 35 fiye da ƙididdiga a cikin cortex na ɗan adam, amma ana buƙatar farashin duka don tsinkayar lambar synapse. Levy, WB & Calvert, VG Sadarwa yana cinye makamashi sau 35 fiye da ƙididdiga a cikin cortex na ɗan adam, amma ana buƙatar farashin duka don tsinkayar lambar synapse.Levy, WB da Calvert, WG Sadarwa yana cinye makamashi sau 35 fiye da ƙididdiga a cikin ƙwayar ɗan adam, amma ana buƙatar farashin duka biyu don tsinkayar adadin synapses. Levy, WB & Calvert, VG Communication 消耗的能量是人类皮层计算的35 Levy, WB & Calvert, Sadarwar VGLevy, WB da Calvert, WG Sadarwa yana cinye makamashi sau 35 fiye da ƙididdigewa a cikin cortex na ɗan adam, amma duka farashin suna buƙatar tsinkayar adadin synapses.tsari. Kwalejin Kimiyya ta Kasa. ilimin kimiyya. US 118, https://doi.org/10.1073/pnas.2008173118 (2021).
Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararru. Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararru.Dalgati, T., Vianello, E., DeSalvo, B. da Casas, J. Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararru.Dalgati T., Vianello E., DeSalvo B. da Casas J. Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararru. A halin yanzu. Ra'ayi. Kimiyyar kwari. 30, 59-66 (2018).
Roy, K., Jaiswal, A. & Panda, P. Zuwa ga na'ura mai tushen karu tare da lissafin neuromorphic. Roy, K., Jaiswal, A. & Panda, P. Zuwa ga na'ura mai tushen karu tare da lissafin neuromorphic. Roy, K., Jaiswal, A. & Panda, P. Zuwa ga Ingantacciyar Injin Tushen Karu tare da Kwamfuta na Neuromorphic.Roy K, Jaiswal A, da Panda P. Pulse na tushen basirar wucin gadi ta amfani da lissafin neuromorphic. Yanayin 575, 607-617 (2019).
Indiveri, G. & Liu, S.-C. Indiveri, G. & Liu, S.-C.Indiveri, G. da Liu, S.-K. Indiveri, G. & Liu, S.-C. Indiveri, G. & Liu, S.-C.Indiveri, G. da Liu, S.-K.Ƙwaƙwalwar ƙwaƙwalwa da sarrafa bayanai a cikin tsarin neuromorphic. tsari. IEEE 103, 1379-1397 (2015).
Akopyan F. et al. Truenorth: Zane da kayan aikin kayan aiki don 65mW 1 miliyan neuron wanda za'a iya tsara guntun synaptic. IEEE ma'amaloli. Tsarin kwamfuta na tsarin da'ira mai haɗaka. 34, 1537-1557 (2015).
Schemmel, J. et al. Live demo: saukar da sigar tsarin neuromorphic na BrainScaleS a sikelin faranti. 2012 IEEE Taro na Kasa da Kasa akan Da'irori da Tsarin (ISCAS), (IEEE ed.) 702-702 (2012).
Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. Tsarin gine-ginen multicore mai daidaitawa tare da tsarin ƙwaƙwalwar ajiya iri-iri don masu sarrafa asynchronous neuromorphic (DYNAPs). Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. Tsarin gine-ginen multicore mai daidaitawa tare da tsarin ƙwaƙwalwar ajiya iri-iri don masu sarrafa asynchronous neuromorphic (DYNAPs).Moradi S., Qiao N., Stefanini F. da Indiviri G. Tsarin gine-ginen multicore mai daidaitawa tare da tsarin ƙwaƙwalwar ajiya iri-iri don masu sarrafa asynchronous neuromorphic (DYNAP). Moradi, S., Qiao, N.、Stefanini, F. & Indiveri, G.内存结构. Moradi, S., Qiao, N.,Stefanini, F. & Indiveri, G. Wani nau'i na gine-gine mai fa'ida da yawa, tare da tsarin ƙwaƙwalwar ajiya na musamman don sarrafa jijiya mai ƙarfi (DYNAP).Moradi S., Qiao N., Stefanini F. da Indiviri G. Tsarin gine-ginen multicore mai daidaitawa tare da tsarin ƙwaƙwalwar ajiya iri-iri don masu sarrafa asynchronous neuromorphic (DYNAP).IEEE Ma'amaloli akan Kimiyyar Halittu. tsarin lantarki. 12, 106-122 (2018).
Davis, M. et al. Loihi: Neuromorphic Multi-core processor tare da ilmantarwa. IEEE Micro 38, 82-99 (2018).
Furber, SB, Galluppi, F., Temple, S. & Plana, LA Aikin SpinNaker. Furber, SB, Galluppi, F., Temple, S. & Plana, LA Aikin SpinNaker.Ferber SB, Galluppi F., Temple S. da shirin Plana LA SpinNnaker.Ferber SB, Galluppi F., Temple S. da shirin Plana LA SpinNnaker. tsari. IEEE 102, 652-665 (2014).
Liu, S.-K. & Delbruck, T. Neuromorphic tsarin jijiya. & Delbruck, T. Neuromorphic tsarin jijiya.da Delbrück T. Neuromorphic tsarin jijiya. & Delbruck, T. 神经形态感觉系统。 & Delbruck, T.da Delbrück T. Neuromorphic tsarin jijiya.A halin yanzu. Ra'ayi. Neurobiology. 20, 288-295 (2010).
Chope, T. et al. Haɗe-haɗe na haɗe-haɗe na neuromorphic don haɗakar da tushen sauti da kuma guje wa karo. A cikin 2019 a taron IEEE akan Matsaloli da Tsarin Halittu (BioCAS), (IEEE Ed.) 1-4 (2019).
Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G. Gine-gine na tushen neuromorphic na hangen nesa na sitiriyo. Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G. Gine-gine na tushen neuromorphic na hangen nesa na sitiriyo.Risi N, Aymar A, Donati E, Solinas S, da Indiveri G. Gine-ginen sitiriyo na tushen karu. Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G. 一种基于脉冲的立体视觉神经形态结构。 Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G.Risi N, Aimar A, Donati E, Solinas S, da Indiveri G. Gine-ginen neuromorphic na tushen Spike don hangen nesa na sitiriyo.gaba. Neurobotics 14, 93 (2020).
Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. Tsarin hanyar sadarwa na 3Dperception don tsarin hangen nesa na sitiriyo na tushen neuromorphic. Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. Tsarin hanyar sadarwa na 3Dperception don tsarin hangen nesa na sitiriyo na tushen neuromorphic.Oswald, M., Ieng, S.-H., Benosman, R., da Indiveri, G. A 3D Pulsed Neural Network Perception Model don Tsarin Neuromorphic Stereo Vision Vision Systems. Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. 3Dperception 脉冲神经网络模型。Oswald, M., Ieng, S.-H., Benosman, R., da Indiveri, G. Spiked 3Dperception Neural Network Model don Tsarin Neuromorphic Neuromorphic Vision Vision System.ilimin kimiyya. Rahoton 7, 1-11 (2017).
Dalgaty, T. et al. Ganewar motsi na asali na kwari ya haɗa da ƙwaƙwalwar juriya da fashewar cibiyoyin sadarwa. Bionic tsarin biohybrid. 10928, 115-128 (2018).
D'Angelo, G. et al. Gano motsi na tushen abubuwan da suka faru ta amfani da ƙididdigewa na ɗan lokaci. gaba. Neurology. 14, 451 (2020).
Lokacin aikawa: Nuwamba-17-2022