Izindaba

Izinhlelo zokusebenza zokucubungula idatha yomhlaba wangempela zidinga amasistimu ekhompuyutha ahlangene, aphansi, anamandla aphansi. Ngamakhono ekhompuyutha aqhutshwa umcimbi, izakhiwo ezihambisanayo ze-metal-oxide-semiconductor hybrid memristive neuromorphic zinikeza isisekelo sehadiwe esifanelekile semisebenzi enjalo. Ukuze sibonise amandla aphelele amasistimu anjalo, siphakamisa futhi sibonise ngokuhlola isisombululo esibanzi sokucubungula inzwa sezinhlelo zokusebenza zokwenziwa kwasendaweni kwezinto zangempela. Drawing inspiration from barn owl neuroanatomy, we have developed a bioinspired, event-driven object localization system that combines a state-of-the-art piezoelectric micromechanical transducer transducer with computational graph-based neuromorphic resistive memory. We show measurements of a fabricated system that includes a memory-based resistive coincidence detector, delay line circuitry, and a fully customizable ultrasonic transducer. Sisebenzisa le miphumela yokuhlola ukuze silinganise izifaniso ezingeni lesistimu. Lezi zifaniso zibe sezisetshenziselwa ukuhlola ukulungiswa kwe-angular nokusebenza kahle kwamandla kwemodeli yokwenza kwasendaweni into. Imiphumela ibonisa ukuthi indlela yethu ingaba ama-oda amaningana wobukhulu bayonga amandla amaningi kunama-microcontroller enza umsebenzi ofanayo.

Kuleli phepha, sethula isistimu yokucubungula yokuzwa eqhutshwa umcimbi esanda kuthuthukiswa esetshenziswa ekwenziweni kwasendaweni kwento. Lapha, ngokokuqala ngqa, sichaza isistimu yokuphela ukuya-ekupheleni yokwenza kwasendaweni kwento etholwe ngokuxhuma i-piezoelectric micromachined ultrasonic transducer (pMUT) yesimanjemanje ngegrafu yekhompyutha esekelwe kumemori ye-neuromorphic resistive (RRAM). I-in-memory computing architectures esebenzisa i-RRAM iyisixazululo esithembisayo sokunciphisa ukusetshenziswa kwamandla23,24,25,26,27,28,29. Ukungaguquguquki kwazo okungokwemvelo—okungadingi ukusetshenziswa kwamandla okusebenzayo ukuze kugcinwe noma kuthuthukiswe ulwazi—kufanelana ngokuphelele nemvelo engavumelaniyo, eqhutshwa umcimbi yekhompyutha ye-neuromorphic, okuholela ekucisheni kungabikho ukusetshenziswa kwamandla lapho isistimu ingenzi lutho. I-Piezoelectric micromachined ultrasonic transducers (pMUTs) ayibizi, ama-miniaturized silicon-based ultrasonic transducers akwazi ukusebenza njengabathumeli nabamukeli30,31,32,33,34. Ukuze sicubungule amasiginali atholwe izinzwa ezakhelwe ngaphakathi, sithole ugqozi kusuka ku-barn owl neuroanatomy35,36,37. Isikhova esibizwa ngokuthi i-barn owl i-Tyto alba saziwa ngamakhono aso amangalisayo okuzingela ebusuku ngenxa yesistimu yokuzwa esebenza kahle kakhulu. Ukuze ubale indawo yesilwane esizingelayo, uhlelo lwendawo yesikhova esibayeni lufaka ikhodi isikhathi sokundiza (i-ToF) lapho amaza omsindo avela enyamazaneni efinyelela ezindlebeni zesikhova ngasinye noma izimukeli zomsindo. Uma kubhekwa ibanga eliphakathi kwezindlebe, umehluko phakathi kwezilinganiso ezimbili ze-ToF (I-Interural Time Difference, i-ITD) yenza kube nokwenzeka ukubala ngokuhlaziya indawo ye-azimuth yethagethi. Although biological systems are poorly suited to solving algebraic equations, they can solve localization problems very effectively. Uhlelo lwemizwa yesikhova esibayeni lusebenzisa isethi yomtshina wokuqondana (CD)35 neurons (okungukuthi, ama-neuron akwazi ukubona ukuhlobana kwesikhashana phakathi kwama-spikes asabalalisa phansi kuya eziphethweni ezijabulisayo eziguqukayo)38,39 ehlelwe yaba amagrafu ekhompyutha ukuze axazulule izinkinga zokuma.
Ucwaningo lwangaphambilini lubonise ukuthi i-complementary metal-oxide-semiconductor (CMOS) hardware kanye ne-RRAM-based neuromorphic hardware ephefumulelwe i-colliculus ephansi (“i-auditory cortex”) yesikhova esibayeni iyindlela ephumelelayo yokubala indawo usebenzisa i-ITD13, 40, 41, 42, 43, 44, 45, 46. Kodwa-ke, amandla ezinhlelo eziphelele ze-neuromorphic ezixhumanisa izinkomba zokuzwa kumagrafu we-neuromorphic computational ayikaboniswa. Inkinga enkulu ukuhlukahluka okungokwemvelo kwamasekhethi e-analog CMOS, okuthinta ukunemba kokutholwa komdlalo. Recently, alternative numerical implementations of the ITD47 estimates have been demonstrated. Kuleli phepha, siphakamisa ukusebenzisa ikhono le-RRAM ukushintsha inani lokuqhuba ngendlela engaguquguquki ukuze kuliwe nokuhlukahluka kumasekhethi e-analog. Senze isistimu yokuhlola ehlanganisa ulwelwesi olulodwa oludlulisa i-pMUT olusebenza ngemvamisa engu-111.9 kHz, i-pMUT emibili ethola ulwelwesi (izinzwa) ezilingisa izindlebe zezikhova ezibayeni, kanye neyodwa . Sihlole ngokuhlola isistimu yokuthola i-pMUT kanye negrafu yekhompyutha ye-ITD esekwe ku-RRAM ukuze sihlole isistimu yethu yokwenza okwasendaweni futhi sihlole ukulungiswa kwayo kwe-angular.
Siqhathanisa indlela yethu nokusetshenziswa kwedijithali ku-microcontroller eyenza umsebenzi ofanayo wokwenza indawo kusetshenziswa izindlela ezivamile ze-beamforming noma ze-neuromorphic, kanye nohlelo lwesango oluhlelekayo lwenkambu (FPGA) lokulinganisa kwe-ITD okuhlongoziwe kusithenjwa. 47. This comparison highlights the competitive power efficiency of the proposed RRAM-based analog neuromorphic system.

isikhova sasesibayeni sithola amaza omsindo avela endaweni eqondiwe, kulokhu sinyakazisa inyamazane. Isikhathi sokundiza (ToF) segagasi lomsindo sihlukile endlebeni ngayinye (ngaphandle uma inyamazane iphambi kwesikhova ngqo). Umugqa onamachashazi ubonisa indlela amaza omsindo ayihambayo ukuze afinyelele ezindlebeni zesikhova esisesibayeni. Inyamazane ingenziwa endaweni ngokunembile endizeni evundlile ngokusekelwe kumehluko wobude phakathi kwezindlela ezimbili ze-acoustic kanye nomehluko ohambisanayo wesikhathi se-interaural (ITD) (isithombe esingakwesokunxele sikhuthazwe yi-ref. 74, copyright 2002, Society for Neuroscience). In our system, the pMUT transmitter (dark blue) generates sound waves that bounce off the target. Amagagasi e-ultrasound abonakalisiwe atholwa izamukeli ezimbili ze-pMUT (okuluhlaza okotshani) futhi acutshungulwe iphrosesa ye-neuromorphic (kwesokudla). b Imodeli yekhompyutha ye-ITD (Jeffress) echaza ukuthi imisindo engena ezindlebeni zesikhova esisesibayeni ifakwa kanjani ikhodi kuqala njengama-spikes akhiywe ngesigaba ku-nucleus enkulu (NM) bese kusetshenziswa igridi ehlelwe ngokwejometri yama-neurons omtshina afanayo kunucleus ye-lamellar. Processing (Netherlands) (left). Illustration of a neuroITD computational graph combining delay lines and coincidence detector neurons, the owl biosensor system can be modeled using RRAM-based neuromorphic circuits (right). c Schematic of the main Jeffress mechanism, due to the difference in ToF, the two ears receive sound stimuli at different times and send axons from both ends to the detector. The axons are part of a series of coincidence detector (CD) neurons, each of which responds selectively to strongly time-correlated inputs. As a result, only CDs whose inputs arrive with the smallest time difference are maximally excited (ITD is exactly compensated). I-CD izobe isifaka ikhodi indawo eyi-angular yalapho okuqondiwe khona.

a Image of a pMUT crystal with six 880 µm membranes integrated at 1.5 mm pitch. b Umdwebo wokusethwa kokulinganisa. Ithagethi itholakala endaweni ye-azimuth θ futhi ebangeni D. Isidluliseli se-pMUT sikhiqiza isignali engu-117.6 kHz egxuma isuke lapho kuqondiwe futhi ifinyelele izamukeli ezimbili ze-pMUT ezinesikhathi esihlukile sokundiza (ToF). Lo mehluko, ochazwa njengomehluko wesikhathi esiphakathi kokuzwakalayo (ITD), ubhala ngekhodi indawo yento futhi ungalinganiselwa ngokulinganisa impendulo ephezulu yezinzwa ezimbili ezamukelayo. c Schematic of pre-processing steps for converting the raw pMUT signal into spike sequences (ie input to the neuromorphic computation graph). Izinzwa ze-pMUT namagrafu ekhompiyutha e-neuromorphic enziwe futhi ahlolwa, futhi ukucubungula kwangaphambili kwe-neuromorphic kusekelwe ekulingiseni isofthiwe. d Response of the pMUT membrane upon receipt of a signal and its transformation into a spike domain. e Ukwenziwa kwasendaweni kokuhlola ukunemba kwe-angular njengomsebenzi we-engeli yento (Θ) nebanga (D) ukuya entweni eqondiwe. Indlela yokukhipha i-ITD idinga ukulungiswa kwe-angular okungenani okungaba ngu-4°C. f Ukunemba kwe-angular (umugqa oluhlaza okwesibhakabhaka) nesilinganiso esihambelanayo se-peak-to-noise (umugqa oluhlaza) uma kuqhathaniswa nebanga lento ye-Θ = 0.
Resistive memory stores information in a non-volatile conductive state. The basic principle of the method is that the modification of the material at the atomic level causes a change in its electrical conductivity57. Lapha sisebenzisa inkumbulo ephikisayo esekwe ku-oxide ehlanganisa isendlalelo esingu-5nm se-hafnium dioxide ehlanganiswe phakathi kwe-titanium engaphezulu nangaphansi kanye nama-electrode e-titanium nitride. I-conductivity yamadivayisi e-RRAM ingashintshwa ngokusebenzisa i-waveform yamanje/i-voltage edala noma enqamule imicu yokuqhuba yezikhala ze-oxygen phakathi kwama-electrode. Sahlanganisa amadivaysi anjalo58 kunqubo evamile engu-130 nm ye-CMOS ukuze sakhe isifunda se-neuromorphic esilungiseka kabusha sisebenzisa umtshina wokuziqondana kanye nesekhethi yomugqa wokulibaziseka (Fig. 3a). Imvelo engaguquki ne-analog yedivayisi, kuhlanganiswe nemvelo eqhutshwa umcimbi yesekethe ye-neuromorphic, inciphisa ukusetshenziswa kwamandla. Isekhethi inomsebenzi wokuvula/ukuvala ngokushesha: isebenza ngokushesha ngemva kokuvulwa, ivumela amandla ukuthi acishwe ngokuphelele lapho isifunda singasebenzi. The main building blocks of the proposed scheme are shown in fig. 3b. Iqukethe izinhlaka ze-N parallel single-resistor single-transistor (1T1R) ezifaka izisindo ze-synaptic lapho kuthathwa khona imisinga enesisindo, ijovwe ku-synapse evamile ye-different pair integrator (DPI)59, futhi ekugcineni ijovwe ku-synapse ngokuhlanganisa kanye leakage. activated (LIF) neuron 60 (see Methods for details). Ukufakwa kokufakwayo kusetshenziswa esangweni lesakhiwo se-1T1R ngendlela yokulandelana kwama-voltage pulses anobude besikhathi ngokulandelana kwamakhulu ama-nanoseconds. Inkumbulo ekwazi ukumelana nayo ingafakwa esimweni sokuqhuba kahle (i-HCS) ngokusebenzisa ireferensi yangaphandle evumayo ku-Vtop lapho i-Vbottom igxiliwe, futhi isethelwe esimweni sokuqhuba esiphansi (LCS) ngokusebenzisa i-voltage ephozithivu ku-Vbottom lapho i-Vtop imisiwe. Isilinganiso senani le-HCS lingalawulwa ngokukhawulela ukusebenza kwamanje (ukuthobela) kwe-SET (ICC) ngevolthi yesango yomthombo we-transistor yochungechunge (Fig. 3c). The functions of RRAM in the circuit are twofold: they direct and weight the input pulses.
Isithombe se-electron microscope (SEM) yeskena sedivayisi ye-HfO2 1T1R RRAM eluhlaza okwesibhakabhaka ehlanganiswe kubuchwepheshe be-CMOS obungu-130 nm nama-transistors esikhethi (650 nm ububanzi) ngokuluhlaza. b Amabhlogo wokwakha ayisisekelo we-schema ye-neuromorphic ehlongozwayo. Ama-voltage okufakwayo ama-pulses (iziqongo) i-Vin0 ne-Vin1 zisebenzisa i-Iweight yamanje, elingana nezifunda zokuqhuba i-G0 ne-G1 yesakhiwo se-1T1R. Lokhu kwamanje kujovwa kuma-synapse e-DPI futhi kujabulise ama-neuron e-LIF. I-RRAM G0 ne-G1 ifakwe ku-HCS ne-LCS ngokulandelanayo. c Umsebenzi we-cumulative conductance density weqembu lamadivayisi we-16K RRAM njengomsebenzi wokufanisa wamanje we-ICC, olawula ngempumelelo izinga lokuqhuba. d Izilinganiso zesiyingi ku-(a) ezibonisa ukuthi i-G1 (ku-LCS) ivimba ngempumelelo okokufaka okuvela ku-Vin1 (okuluhlaza), futhi ngempela i-voltage ye-neuron ephumayo isabela kuphela kokokufaka okuluhlaza okwesibhakabhaka okuvela ku-Vin0. RRAM effectively determines the connections in the circuit. e Ukulinganiswa kwesekhethi ku-(b) okubonisa umthelela wevelu ye-conductance G0 ku-voltage yolwelwesi lwe-Vmem ngemva kokufaka i-voltage pulse Vin0. Ukuziphatha okwengeziwe, impendulo inamandla: ngakho-ke, idivayisi ye-RRAM isebenzisa isisindo sokuxhuma se-I/O. Izilinganiso zenziwe kusekethe futhi zikhombisa ukusebenza okumbaxambili kwe-RRAM, umzila kanye nesisindo sama-pulses okokufaka.
Okokuqala, njengoba kunezimo ezimbili eziyisisekelo zokuqhuba (i-HCS ne-LCS), ama-RRAM angavimba noma aphuthe ama-pulses okokufaka uma esezifundeni ze-LCS noma ze-HCS, ngokulandelanayo. As a result, RRAM effectively determines the connections in the circuit. This is the basis for being able to reconfigure the architecture. Ukuze sibonise lokhu, sizochaza ukuqaliswa kwesekethe esunguliwe yebhlokhi yesifunda ku-Fig. 3b. I-RRAM ehambisana ne-G0 ihlelwe ku-HCS, futhi i-RRAM G1 yesibili ihlelwe yaba yi-LCS. Input pulses are applied to both Vin0 and Vin1. The effects of two sequences of input pulses were analyzed in the output neurons by collecting the neuron membrane voltage and the output signal using an oscilloscope. Ukuhlolwa kube yimpumelelo lapho kuphela idivayisi ye-HCS (G0) ixhunywe kukushaya kwemithambo ye-neuron ukuze kushukunyiswe ukungezwani kolwelwesi. This is demonstrated in Figure 3d, where the blue pulse train causes the membrane voltage to build up on the membrane capacitor, while the green pulse train keeps the membrane voltage constant.
The second important function of RRAM is the implementation of connection weights. Using RRAM's analog conductance adjustment, I/O connections can be weighted accordingly. In the second experiment, the G0 device was programmed to different levels of HCS, and the input pulse was applied to the VIn0 input. I-pulse yokufaka idonsa i-current (Iweight) kusukela ocingweni, elingana nendlela yokuqhuba kanye ne-Vtop okungenzeka ibe khona ehambisanayo − Vbot. This weighted current is then injected into the DPI synapses and LIF output neurons. The membrane voltage of the output neurons was recorded using an oscilloscope and displayed in Fig. 3d. The voltage peak of the neuron membrane in response to a single input pulse is proportional to the conductance of the resistive memory, demonstrating that RRAM can be used as a programmable element of synaptic weight. Lezi zivivinyo ezimbili zokuqala zibonisa ukuthi iplathifomu ehlongozwayo ye-neuromorphic esekelwe ku-RRAM iyakwazi ukusebenzisa izici eziyisisekelo zendlela ye-Jeffress eyisisekelo, okungukuthi umugqa wokulibaziseka kanye nesekhethi yomtshina wokuhlangana. Iplathifomu yesifunda yakhiwe ngokunqwabelanisa amabhlogo alandelanayo, njengamabhlogo akuMfanekiso 3b, nokuxhuma amasango awo emugqeni wokufaka ovamile. We designed, fabricated, and tested a neuromorphic platform consisting of two output neurons receiving two inputs (Fig. 4a). Umdwebo wesifunda ukhonjiswe kuMfanekiso 4b. I-matrix ye-2 × 2 RRAM ephezulu ivumela ama-pulses okokufaka ukuthi aqondiswe kuma-neurons amabili okukhiphayo, kuyilapho i-2 × 2 matrix ephansi ivumela ukuxhumeka okuphindaphindiwe kwama-neurons amabili (N0, N1). Sibonisa ukuthi le nkundla ingasetshenziswa ngokucushwa komugqa wokulibaziseka kanye nemisebenzi emibili ehlukene yesitholi sezinto eziqondane, njengoba kuboniswa izilinganiso zokuhlola ku-Fig. 4c-e.
Umdwebo wesiyingi owenziwe ama-neuron amabili okukhiphayo i-N0 kanye ne-N1 ethola okokufaka okubili okungu-0 no-1. Amadivayisi amane aphezulu ohlwini achaza ukuxhumana kwe-synaptic kusuka kokokufakayo kuye kokukhiphayo, futhi amaseli amane aphansi achaza ukuxhumana okuphindelelayo phakathi kwama-neurons. Ama-RRAM anemibala amelela amadivayisi alungiselelwe ku-HCS kwesokudla: amadivayisi ku-HCS avumela uxhumo futhi amele izisindo, kuyilapho amadivayisi aku-LCS avimba ama-pulses okokufaka futhi akhubaza ukuxhumeka kokuphumayo. b Umdwebo wesekethe (a) onamamojula ayisishiyagalombili e-RRAM agqanyiswe ngokuluhlaza okwesibhakabhaka. c Imigqa yokulibaziseka yenziwa ngokusebenzisa nje amandla e-DPI synapses nama-LIF neurons. I-RRAM eluhlaza isethelwe ekuqhubeni phezulu ngokwanele ukuze ikwazi ukubangela iphutha ekuphumeni ngemva kokubambezeleka kokokufaka Δt. d Schematic illustration of direction-insensitive CD detection of time dependent signals. Output neuron 1, N1, fires on inputs 0 and 1 with a short delay. e Isekhethi ye-CD ezwelayo ekuqondeni, isekethe ethola lapho okokufaka 1 kusondela ku-0 futhi ifike ngemva kokufaka 0. Okukhiphayo kwesekhethi kumelelwe yi-neuron 1 (N1).


a Experimental measurements of the response of nine randomly selected individual neurons to the same input pulse. The response varies across populations, affecting input gain and time constant. b Izilinganiso zokuhlola zethonya lama-neuron ekuguquguqukeni kwama-neuron athinta i-CD ezwela ukuqondiswa. Ama-neuron amabili azwelayo e-CD asabela ngendlela ehlukile kusisusa sokufakwayo ngenxa yokuhlukahluka kwe-neuron-to-neuron. Neuron 0 has a lower input gain than neuron 1, so it takes three input pulses (instead of 1) to create an output spike. As expected, neuron 1 reaches the threshold with two input events. If input 1 arrives Δt = 50 µs after neuron 0 fires, CD remains silent because Δt is greater than the time constant of neuron 1 (about 22 µs). c is reduced by Δt = 20 µs, so that input 1 peaks when neuron 1′s firing is still high, resulting in the simultaneous detection of two input events.
Izinto ezimbili ezisetshenziswe kukholomu yokubala ye-ITD umugqa wokulibaziseka kanye ne-CD enganaki isiqondisindlela. Womabili amasekhethi adinga ukulinganisa okunembile ukuze kuqinisekiswe ukusebenza kahle kokuma kwento. Ulayini wokulibaziseka kufanele ulethe inguqulo ebambezeleke ngokunembile yokuphakama kokufaka (Fig. 6a), futhi i-CD kufanele yenziwe isebenze kuphela uma okokufaka kuwela phakathi kwebanga lokuthola okuqondiwe. Ngomugqa wokulibaziseka, izisindo ze-synaptic zokuxhuma okokufaka (i-G3 ku-Fig. 4a) zahlelwa kabusha kuze kutholakale ukubambezeleka okuhlosiwe. Setha ukubekezelela eduze nokulibaziseka okuqondiwe ukuze umise uhlelo: uma kuncane ukubekezelelana, kuba nzima kakhulu ukusetha ngempumelelo umugqa wokulibaziseka. Emkhiwaneni. Umfanekiso 6b ubonisa imiphumela yenqubo yokulinganisa umugqa wokulibaziseka: kungabonakala ukuthi isikimu esihlongozwayo singanikeza ngokunembile konke ukubambezeleka okudingekayo esikimini sokuklama (kusuka ku-10 kuya ku-300 μs). Inombolo enkulu yokuphindaphinda kokulinganisa ithinta ikhwalithi yenqubo yokulinganisa: ukuphindaphinda okungu-200 kunganciphisa iphutha libe ngaphansi kuka-5%. Ukulinganisa okukodwa kuhambisana nokusebenza kokusetha/ukusetha kabusha kweseli ye-RRAM. Inqubo yokushuna ibalulekile futhi ekuthuthukiseni ukunemba kwemojuli ye-CD yokutholwa komcimbi oseduze ngokushesha. Kuthathe izikhathi eziyishumi zokulinganisa ukulinganisa ukuze kuzuzwe izinga leqiniso (okungukuthi, izinga lezehlakalo ezihlonzwe ngendlela efanele njengezifanele) ngaphezu kwama-95% (umugqa oluhlaza kuFigure 6c). Kodwa-ke, inqubo yokushuna ayizange ithinte izehlakalo ezinhle ezingamanga (okungukuthi, imvamisa yemicimbi ekhonjwe ngephutha njengezifanele). Enye indlela ebonwa ezinhlelweni zebhayoloji yokunqoba izithiyo zesikhathi sezindlela ezisebenza ngokushesha ukuphindaphinda (okungukuthi, amakhophi amaningi ento efanayo asetshenziselwa ukwenza umsebenzi othile). Sikhuthazwe yi-biology66, sibeke amasekhethi ama-CD amaningana kumojula ngayinye ye-CD phakathi kwemigqa yokulibaziseka emibili ukuze sinciphise umthelela wezinto ezingamanga. Njengoba kuboniswe emkhiwaneni. 6c (umugqa oluhlaza), ukubeka izici ze-CD ezintathu kumojula ngayinye ye-CD kungehlisa izinga le-alamu elingamanga libe ngaphansi kuka-10–2.
Umthelela wokuhlukahluka kwe-neuronal kumasekhethi omugqa wokulibaziseka. b Amasekhethi omugqa wokulibaziseka angalinganiswa abe ukubambezeleka okukhulu ngokusetha izikhathi ezingaguquki zama-neurons e-LIF ahambisanayo nama-synapse e-DPI kube amanani amakhulu. Ukwenyusa inani lokuphindaphinda kwenqubo yokulinganisa ye-RRAM kwenze kwaba nokwenzeka ukuthuthukisa ngokuphawulekayo ukunemba kokulibaziseka okuqondiwe: ukuphindaphinda okungu-200 kwehlise iphutha laba ngaphansi kuka-5%. Ukuphindaphinda okukodwa kuhambisana nokusebenza kwe-SET/RESET kuseli ye-RRAM. Imojula ngayinye ye-CD kumodeli ye-c Jeffress ingasetshenziswa kusetshenziswa ama-elementi e-CD ahambisanayo angu-N ukuze kube nokuguquguquka okukhulu ngokuphathelene nokuhluleka kwesistimu. d Ukulinganisa okwengeziwe kwe-RRAM kokulinganisa kwenyusa izinga leqiniso eliphozithivu (umugqa oluhlaza), kuyilapho izinga elingelona iqiniso lizimele kwinani lokuphindaphinda (umugqa oluhlaza). Ukubeka izinto eziningi ze-CD ngokuhambisana kugwema ukutholwa okungamanga kokufana kwamamojula e-CD.

ukulungiswa kwe-Angular (okuluhlaza okwesibhakabhaka) kanye nokusetshenziswa kwamandla (okuluhlaza) kokusebenza kokwenziwa kwasendaweni kuye ngenani lamamojula e-CD. Ibha evundlile eluhlaza okwesibhakabhaka okumnyama imelela ukunemba kwe-angular kwe-PMUT kanye nebha evundlile eluhlaza okwesibhakabhaka ngokukhanyayo imelela ukunemba kwe-angular kwegrafu yekhompyutha ye-neuromorphic. b Ukusetshenziswa kwamandla kwesistimu ehlongozwayo kanye nokuqhathanisa nokusetshenziswa kokubili okuxoxiwe kwe-microcontroller kanye nokuqaliswa kwedijithali kwe-Time Difference Encoder (TDE)47 FPGA.
Ukuze kuncishiswe ukusetshenziswa kwamandla kwesistimu yendawo eqondiwe, sicabange, saklama futhi sasebenzisa isekethe ye-neuromorphic esekelwe ku-RRAM eqhutshwa umcimbi esebenza kahle ecubungula ulwazi lwesiginali olukhiqizwe izinzwa ezakhelwe ngaphakathi ukuze sibale indawo yento eqondiwe ngokoqobo. isikhathi. . Nakuba izindlela zokucubungula zendabuko ziqhubeka nokusampula amasignali atholiwe futhi zenza izibalo ukuze kukhishwe ulwazi oluwusizo, isixazululo esihlongozwayo se-neuromorphic senza izibalo ngendlela efanayo njengoba ulwazi oluwusizo lufika, lukhulise ukusebenza kahle kwamandla esistimu ngama-oda amahlanu obukhulu. Ngaphezu kwalokho, sigqamisa ukuguquguquka kwamasekethe e-neuromorphic asuselwa ku-RRAM. Ikhono le-RRAM lokushintsha ukusebenza ngendlela engaguquguquki (i-plasticity) linxephezela ukuhlukahluka okungokwemvelo kwe-ultra-low power analog DPI's synaptic and neural circuits. Lokhu kwenza le sekethi esekwe ku-RRAM ibe nezinto eziningi futhi ibe namandla. Umgomo wethu awukona ukukhipha imisebenzi eyinkimbinkimbi noma amaphethini kumasiginali, kodwa ukwenza izinto zibe ezasendaweni ngesikhathi sangempela. Isistimu yethu ingacindezela kahle isiginali futhi ekugcineni iyithumele ezinyathelweni eziqhubekayo zokucubungula ukwenza izinqumo eziyinkimbinkimbi uma kudingeka. Kumongo wezinhlelo zokusebenza zasendaweni, isinyathelo sethu sokucubungula kusengaphambili se-neuromorphic singanikeza ulwazi mayelana nendawo yezinto. Lolu lwazi lungasetshenziswa, isibonelo, ukuthola ukunyakaza noma ukubonwa kokuthinta. We emphasize the importance of combining ultra low power sensors such as pMUTs with ultra low power electronics. Kulokhu, izindlela ze-neuromorphic zibe ukhiye njengoba zisiholele ekuthuthukiseni ukusetshenziswa kwesekethe okusha kwezindlela zekhompiyutha eziphefumulelwe ngokwebhayoloji njengemodeli ye-Jeffress. Ngomongo wezinhlelo zokusebenza zokuhlanganisa izinzwa, isistimu yethu ingahlanganiswa nezinzwa eziningana ezisekelwe kumcimbi ukuze kutholwe ulwazi olunembe kakhudlwana. Nakuba izikhova zizinhle kakhulu ekutholeni inyamazane ebumnyameni, zikwazi ukubona kahle kakhulu futhi zenza ukusesha okuhlangene kokuzwa nokubukwayo ngaphambi kokubamba inyamazane70. Lapho i-neuron ethile ezwayo ivutha umlilo, isikhova sithola ulwazi esiludingayo ukuze sinqume ukuthi singaqala ngakuphi ukusesha kwaso okubukwayo, ngaleyo ndlela sigxilise ukunaka kwaso engxenyeni encane yendawo yokubuka. Inhlanganisela yezinzwa ezibukwayo (ikhamera ye-DVS) kanye nenzwa yokulalela ehlongozwayo (ngokusekelwe ku-pMUT) kufanele ihlolwe ukuze kuthuthukiswe ama-agent azayo azimele.
Inzwa ye-pMUT itholakala ku-PCB enazamukeli ezimbili eziqhelelene cishe ngo-10 cm, futhi isidlulisi sitholakala phakathi kwabamukeli. Kulo msebenzi, ulwelwesi ngalunye luyisakhiwo se-bimorph esimisiwe esihlanganisa izingqimba ezimbili ze-piezoelectric aluminium nitride (AlN) engu-800 nm obukhulu ehlanganiswe phakathi kwezingqimba ezintathu ze-molybdenum (Mo) engu-200 nm enogqinsi futhi embozwe ngengqimba engu-200 nm obukhulu. isendlalelo se-SiN esiphezulu esidlulayo njengoba kuchazwe kusithenjwa. 71. Ama-electrode angaphakathi nangaphandle asetshenziselwa izingqimba ezingezansi nangaphezulu ze-molybdenum, kuyilapho i-electrode ephakathi ye-molybdenum ingenakulinganiswa futhi isetshenziswe njengomhlabathi, okuholela kulwelwesi olunamapheya amane ama-electrode.
Lesi sakhiwo sivumela ukusetshenziswa kwe-membrane deformation evamile, okuholela ekudluliseni okuthuthukisiwe nokuthola ukuzwela. I-pMUT enjalo ngokuvamile ibonisa ukuzwela kwenjabulo okungu-700 nm/V njengesithumeli, ihlinzeka ngomfutho ongaphezulu ongu-270 Pa/V. Njengomamukeli, ifilimu eyodwa ye-pMUT ibonisa ukuzwela kwesifunda esifushane okungu-15 nA/Pa, okuhlobene ngokuqondile ne-piezoelectric coefficient ye-AlN. Ukuhlukahluka kobuchwepheshe be-voltage kungqimba lwe-AlN kuholela ekushintsheni kwemvamisa ye-resonant, enganxeshezelwa ngokusebenzisa ukuchema kwe-DC ku-pMUT. Ukuzwela kwe-DC kukalwe ku-0.5 kHz/V. Ukuze kufakwe abalingisi be-acoustic, kusetshenziswa imakrofoni ngaphambi kwe-pMUT.
Ukuze silinganise i-echo pulse, sibeke ipuleti elingunxande elinendawo engaba ngu-50 cm2 ngaphambi kwe-pMUT ukukhombisa amaza omsindo akhishiwe. Both the distance between the plates and the angle relative to the pMUT plane are controlled using special holders. Umthombo wamandla kagesi we-Tectronix CPX400DP uchemile ulwelwesi lwe-pMUT oluthathu, lulungisa imvamisa ye-resonant iye ku-111.9 kHz31, kuyilapho ama-transmitters ashayelwa i-Tectronix AFG 3102 pulse generator eshunswe imvamisa ezwakalayo (111.9 kHz) kanye nomjikelezo womsebenzi ongu-0.01. Ama-currents afundwa kumachweba amane okukhiphayo kwesamukeli ngasinye se-pMUT aguqulelwa kuma-voltage kusetshenziswa ukwakheka okukhethekile okuhlukile kwamanje kanye ne-voltage, futhi amasiginali aphumayo afakwa kudijithali uhlelo lokutholwa kwedatha ye-Spektrum. Umkhawulo wokutholwa wawubonakala ngokutholwa kwesignali ye-pMUT ngaphansi kwezimo ezihlukene: sihambise isibonisi kumabanga ahlukene [30, 40, 50, 60, 80, 100] cm futhi sashintsha i-engeli yokusekela ye-pMUT ([0, 20, 40] o ) Figure 2b shows the temporal ITD detection resolution depending on the corresponding angular position in degrees.


Ukusebenzisa amandla e-analog electronics ukulingisa ukuziphatha kwama-neurons nama-synapses kuyisixazululo esihle nesisebenzayo sokuthuthukisa ukusebenza kahle kwekhompyutha. The disadvantage of this computational underlay is that it will vary from scheme to scheme. We quantified the variability of neurons and synaptic circuits (Supplementary Fig. 2a,b). Kukho konke ukubonakaliswa kokuhlukahluka, lokho okuhlotshaniswa nokungaguquguquki kwesikhathi kanye nenzuzo yokokufaka kunomthelela omkhulu kakhulu ezingeni lesistimu. Isikhathi esingaguquguquki se-LIF neuron kanye ne-synapse ye-DPI kunqunywa isekethe ye-RC, lapho inani lika-R lilawulwa amandla kagesi okuchema asetshenziswa esangweni le-transistor (i-Vlk ye-neuron ne-Vtau ye-synapse), inquma ukuvuza izinga. Ukuzuza kokokufaka kuchazwa ngokuthi i-voltage ephakeme kakhulu efinyelelwe yi-synaptic kanye ne-neuronal membrane capacitor evuselelwe ukushaya kwenhliziyo okokufaka. The input gain is controlled by another bias transistor which modulates the input current. Ukulingisa kwe-Monte Carlo okulinganiselwe kunqubo ye-ST Microelectronics' 130nm kwenziwe ukuze kuqoqwe inzuzo ethile yokokufaka kanye nezibalo ezingaguquki zesikhathi. Imiphumela yethulwa Kumfanekiso Owengeziwe wesi-2, lapho inzuzo yokufaka nesikhathi esingaguquki kubalwa njengomsebenzi we-voltage echema elawula izinga lokuvuza. Green markers quantify the standard deviation of the time constant from the mean. Both neurons and synaptic circuits were able to express a wide range of time constants in the range of 10-5-10-2 s, as shown in Supplementary Fig. scheme. Input amplification (Supplementary Fig. 2e,d) of neuronal and synapse variability was approximately 8% and 3%, respectively. Such a deficiency is well documented in the literature: various measurements were performed on the array of DYNAP chips to assess the mismatch between populations of LIF63 neurons. The synapses in the BrainScale mixed signal chip were measured and their inconsistencies analyzed, and a calibration procedure was proposed to reduce the effect of system-level variability64.
Umsebenzi we-RRAM kumasekethe we-neuromorphic ukabili: incazelo yezakhiwo (okufakwayo komzila kokuphumayo) kanye nokuqaliswa kwezisindo ze-synaptic. Impahla yokugcina ingasetshenziselwa ukuxazulula inkinga yokuhlukahluka kwamasekhethi enziwe nge-neuromorphic. Senze inqubo yokulinganisa elula ebandakanya ukuhlela kabusha idivayisi ye-RRAM kuze kube yilapho umjikelezo ohlaziywayo uhlangabezana nezimfuneko ezithile. Okokufaka okunikeziwe, okukhiphayo kuyagadwa futhi i-RRAM ihlelwa kabusha kuze kube kufinyelelwa ukuziphatha okuqondiwe. Isikhathi sokulinda samasekhondi angu-5 safakwa phakathi kwemisebenzi yokuhlela ukuze kuxazululwe inkinga yokuphumula kwe-RRAM okuholela ekuguquguqukeni kokuziphatha kwesikhashana (Ulwazi Olwengeziwe). Izisindo ze-Synaptic ziyalungiswa noma zilinganiswe ngokuya ngezidingo zesekethe ye-neuromorphic emodelayo. Inqubo yokulinganisa ifinyezwa kuma-algorithms engeziwe [1, 2] agxile ezicini ezimbili ezibalulekile zamapulatifomu e-neuromorphic, imigqa yokulibaziseka kanye ne-CD engazweli. Kumjikelezo onomugqa wokulibaziseka, ukuziphatha okuqondiwe ukuhlinzeka nge-pulse ephumayo ngokulibaziseka Δt. Uma ukubambezeleka kwesekethe kwangempela kungaphansi kwevelu eqondiwe, isisindo se-synaptic se-G3 kufanele sehliswe (i-G3 kufanele isethwe kabusha bese isethelwa ku-Icc yamanje efanayo). Ngokuphambene, uma ukubambezeleka kwangempela kukukhulu kunevelu eqondiwe, ukuqhuba kwe-G3 kufanele kukhushulwe (i-G3 kufanele iqale isethwe kabusha bese isethelwa kunani eliphezulu le-Icc). Le nqubo iphindaphindiwe kuze kube yilapho ukubambezeleka okukhiqizwa yisekethe kufana nenani eliqondiwe futhi ukubekezelelana kusethelwe ukumisa inqubo yokulinganisa. Kuma-CD angezwani ne-orientation, amadivayisi amabili e-RRAM, i-G1 ne-G3, ayabandakanyeka enqubweni yokulinganisa. Lo mjikelezo unokufaka okubili, i-Vin0 ne-Vin1, ibambezeleke yi-dt. Isekhethi kufanele iphendule kuphela ekubambezelekeni okungaphansi kobubanzi obufanayo [0,dtCD]. Uma lingekho inani eliphakeme lokukhiphayo, kodwa inani eliphakeme lokokufaka liseduze, womabili amadivayisi e-RRAM kufanele akhuliswe ukuze kusizwe i-neuron ifinyelele embundwini. Ngokuphambene, uma isifunda siphendula ekubambezelekeni okudlula ububanzi obuhlosiwe be-dtCD, ukuqhutshwa kufanele kwehliswe. Repeat the process until the correct behavior is obtained. Compliance current can be modulated by the built-in analog circuit in ref. 72.73. With this built-in circuit, such procedures can be performed periodically to calibrate the system or reuse it for another application.

Idatha esekela imiphumela yalolu cwaningo iyatholakala kumbhali ofanele, i-FM, uma kunesicelo esifanele.
I-Indiveri, G. & Sandamirskaya, Y. Ukubaluleka kwesikhala nesikhathi sokucubungula isignali kuma-ejenti we-neuromorphic: Inselele yokuthuthukisa amandla aphansi, ama-agent azimele asebenzisana nemvelo. I-Indiveri, G. & Sandamirskaya, Y. Ukubaluleka kwesikhala nesikhathi sokucubungula isignali kuma-ejenti we-neuromorphic: Inselele yokuthuthukisa amandla aphansi, ama-agent azimele asebenzisana nemvelo.I-Indiveri G. kanye ne-Sandamirskaya Y. Ukubaluleka kwesikhala nesikhathi sokucubungula isignali kuma-agent we-neuromorphic: inselele yokuthuthukisa ama-agent anamandla aphansi azimele asebenzisana nemvelo. I-Induveri, G. & Sandamirskaya, Y. 空间和时间对于神经形态代理中信号处理的重要性: 开发与环境交互的低功耗, 自主代理的挑战. I-Indiveri, G. & Sandamirskaya, Y.I-Indiveri G. kanye ne-Sandamirskaya Y. Ukubaluleka kwesikhala nesikhathi sokucubungula isignali kuma-agent we-neuromorphic: inselele yokuthuthukisa ama-agent anamandla aphansi azimele asebenzisana nemvelo.IEEE Signal Processing. Ijenali 36, 16–28 (2019).
I-Thorpe, SJ Peak Isikhathi Sokufika: Uhlelo Olusebenzayo Lwekhodi Yenethiwekhi Ye-Neural. e-Eckmiller, R., Hartmann, G. & Hauske, G. (eds). e-Eckmiller, R., Hartmann, G. & Hauske, G. (eds).e-Eckmiller, R., Hartmann, G. and Hauske, G. (eds.).Ku-Eckmiller, R., Hartmann, G., and Hauske, G. (eds.). Ukucubungula okufanayo kumasistimu we-neural namakhompyutha 91-94 (North-Holland Elsevier, 1990).
I-Levy, i-WB & Calvert, i-VG Communication idla amandla aphindwe izikhathi ezingu-35 kune-computing ku-cortex yomuntu, kodwa zombili izindleko ziyadingeka ukubikezela inombolo ye-synapse. I-Levy, i-WB & Calvert, i-VG Communication idla amandla aphindwe izikhathi ezingu-35 kune-computing ku-cortex yomuntu, kodwa zombili izindleko ziyadingeka ukubikezela inombolo ye-synapse.I-Levy, i-WB ne-Calvert, i-WG Communication idla amandla aphindwe izikhathi ezingu-35 kune-computing ku-cortex yomuntu, kodwa zombili izindleko ziyadingeka ukubikezela inani lama-synapses. Levy, WB & Calvert, VG Communication 消耗的能量是人类皮层计算的35 倍，但這两种成本都需要预测突触数量. Levy, WB & Calvert, VG CommunicationI-Levy, i-WB ne-Calvert, i-WG Communication idla amandla aphindwe izikhathi ezingu-35 kune-computing ku-cortex yomuntu, kodwa zombili izindleko zidinga ukubikezela inani lama-synapses.inqubo. I-National Academy of Science. isayensi. I-US 118, https://doi.org/10.1073/pnas.2008173118 (2021).
I-Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. I-insect-inspired neuromorphic computing. I-Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. I-insect-inspired neuromorphic computing.Dalgati T., Vianello E., DeSalvo B. kanye ne-Casas J. I-insect-inspired neuromorphic computing. Current. Opinion. Insect science. 30, 59–66 (2018).
Roy, K., Jaiswal, A. & Panda, P. Towards spike-based machine intelligence ene-neuromorphic computing. Roy, K., Jaiswal, A. & Panda, P. Towards spike-based machine intelligence ene-neuromorphic computing. Roy, K., Jaiswal, A. & Panda, P. Towards Spike-based Machine Intelligence ene-Neuromorphic Computing.U-Roy K, u-Jaiswal A, kanye no-Panda P. Ubuhlakani bokwenziwa obususelwa kuPulse kusetshenziswa i-neuromorphic computing. Imvelo 575, 607–617 (2019).
I-Indiveri, G. & Liu, S.-C. I-Indiveri, G. & Liu, S.-C.I-Indiveri, G. kanye no-Liu, S.-K. I-Indiveri, G. & Liu, S.-C. I-Indiveri, G. & Liu, S.-C.I-Indiveri, G. kanye no-Liu, S.-K.Memory and information processing in neuromorphic systems. inqubo. IEEE 103, 1379–1397 (2015).
U-Akopyan F. et al. Truenorth: Design and toolkit for a 65 mW 1 million neuron programmable synaptic chip. Ukushintsha kwe-IEEE. Ukwakhiwa kwekhompyutha yezinhlelo zesekethe ezididiyelwe. 34, 1537–1557 (2015).
Schemmel, J. et al. Idemo ebukhoma: inguqulo eyehlisiwe yesistimu ye-BrainScaleS neuromorphic esikalini sepuleti. 2012 IEEE International Symposium on Circuits and Systems (ISCAS), (IEEE ed.) 702–702 (2012).
I-Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. I-architecture ye-multicore scalable enezakhiwo zememori ezihlukile zama-dynamic neuromorphic asynchronous processors (DYNAPs). I-Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. I-architecture ye-multicore scalable enezakhiwo zememori ezihlukile zama-dynamic neuromorphic asynchronous processors (DYNAPs).I-Moradi S., Qiao N., Stefanini F. kanye ne-Indiviri G. I-architecture ye-multicore ekalakayo enezakhiwo zememori ezihlukile zama-dynamic neuromorphic asynchronous processors (DYNAP). Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. 一种可扩展的多核架构，具有用于动态神经形态异步处理器(DYNAP) I-Moradi, S., Qiao, N., iStefanini, F. & Indiveri, G. Uhlobo olunwebekayo lwezakhiwo eziningi eziyisisekelo, ezinesakhiwo sememori esiyingqayizivele sokucubungula kwe-neural (DYNAP).I-Moradi S., Qiao N., Stefanini F. kanye ne-Indiviri G. I-architecture ye-multicore ekalakayo enezakhiwo zememori ezihlukile zama-dynamic neuromorphic asynchronous processors (DYNAP).IEEE Transactions on Biomedical science. electrical system. 12, 106–122 (2018).
UDavis, M. et al. I-Loihi: I-neuromorphic multi-core processor enokufunda okushunyekiwe. IEEE Micro 38, 82–99 (2018).
I-Furber, SB, Galluppi, F., Temple, S. & Plana, LA Iphrojekthi ye-SpiNNaker.Ferber SB, Galluppi F., Temple S. kanye nephrojekthi ye-Plana LA SpiNNaker.Ferber SB, Galluppi F., Temple S. kanye nephrojekthi ye-Plana LA SpiNNaker. inqubo. IEEE 102, 652–665 (2014).
U-Liu, S.-K. & Delbruck, T. Neuromorphic sensory systems. & Delbruck, T. Neuromorphic sensory systems.kanye ne-Delbrück T. Neuromorphic sensory systems. & Delbruck, T. 神经形态感觉系统. & Delbruck, T.kanye ne-Delbrück T. Neuromorphic sensory system.Okwamanje. Umbono. I-Neurobiology. 20, 288–295 (2010).
Chope, T. et al. Ukuhlanganiswa kwezinzwa ze-Neuromorphic zokwenziwa kwasendaweni komthombo womsindo ohlanganisiwe nokugwema ukungqubuzana. Ngo-2019 eNgqungqutheleni ye-IEEE ye-Biomedical Circuits and Systems (BioCAS), (IEEE Ed.) 1–4 (2019).
Risi N, Aymar A, Donati E, Solinas S, kanye ne-Indiveri G. I-spike-based neuromorphic stereovision architecture. 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, kanye ne-Indiveri G. Spike-based neuromorphic architecture yombono we-stereo.ngaphambili. I-Neurorobotics 14, 93 (2020).
Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. Imodeli yenethiwekhi ye-spiking neural ye-3Dperception yezinhlelo zombono we-neuromorphic stereo. Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. Imodeli yenethiwekhi ye-spiking neural ye-3Dperception yezinhlelo zombono we-neuromorphic stereo.Oswald, M., Ieng, S.-H., Benosman, R., kanye ne-Indiveri, G. A 3D Pulsed Neural Network Perception Model for Event-Based Neuromorphic Stereo Vision Systems. Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. 基于事件的神经形态立体视觉系统的3Dperception 脉冲神经网统模。 U-Oswald, M., Ieng, S.-H., Benosman, R., kanye no-Indiveri, G. Spiked 3Dperception Neural Network Model ye-Event-based Neuromorphic Stereo Vision System.

D'Angelo, G. et al. Ukutholwa kokunyakaza okusekelwe kumcimbi kusetshenziswa ikhodi yokuhlukanisa yesikhashana. ngaphambili. I-Neurology. 14, 451 (2020).