Litaba

Lisebelisoa tsa 'nete tsa lefats'e tsa ts'ebetso ea data li hloka litsamaiso tsa komporo tsa komporo, tse tlase-tlase, tse nang le matla a tlase. Ka bokhoni ba likhomphutha tse tsamaisoang ke ketsahalo, meaho e tlatselletsang ea metal-oxide-semiconductor hybrid memristive neuromorphic e fana ka motheo o nepahetseng oa lisebelisoa bakeng sa mesebetsi e joalo. Ho bonts'a bokhoni bo felletseng ba litsamaiso tse joalo, re sisinya le ho bonts'a ka liteko tharollo e felletseng ea ts'ebetso ea sensor bakeng sa lits'ebetso tsa lefats'e tsa nnete. Ho latela ts'usumetso ho tsoa ho barn owl neuroanatomy, re thehile "bioinspired", e tsamaisoang ke liketsahalo sebakeng sa sistimi e kopanyang transducer ea boemo bo holimo ea piezoelectric micromechanical transducer le computational graph-based neuromorphic resistive memory. Re bonts'a litekanyo tsa sistimi e iqapetsoeng e kenyelletsang mochini oa ho tseba lintho tse etsahetseng mohopolong, ho lieha ho potoloha, le transducer ea ultrasonic e ka khonehang ka botlalo. Re sebelisa liphetho tsena tsa liteko ho lekanya lipapiso boemong ba sistimi. Lipapiso tsena li sebelisoa ho lekola qeto ea angular le matla a matla a mofuta oa sebaka sa ntho. Liphetho li bonts'a hore mokhoa oa rona e ka ba litaelo tse 'maloa tsa matla a matla ho feta li-microcontrollers tse etsang mosebetsi o tšoanang.
Re kena mehleng ea likhomphutha tse fumanehang hohle moo palo ea lisebelisoa le lits'ebetso tse sebelisoang e ntseng e hola ka mokhoa o makatsang ho re thusa bophelong ba rona ba letsatsi le letsatsi. Litsamaiso tsena li lebeletsoe hore li sebetse ka ho tsoelang pele, li ja matla a fokolang ka hohle kamoo ho ka khonehang ha li ntse li ithuta ho hlalosa lintlha tseo li li bokellang ho tsoa ho li-sensor tse ngata ka nako ea sebele le ho hlahisa tlhahiso ea binary ka lebaka la ho arola kapa mesebetsi ea ho lemoha. E 'ngoe ea mehato ea bohlokoa ka ho fetisisa e hlokahalang ho finyella pakane ena ke ho ntša boitsebiso bo molemo le bo kopanetsoeng ho tloha boitsebisong bo lerata le hangata bo sa fellang. Mekhoa e tloaelehileng ea boenjiniere e atisa ho etsa mohlala oa matšoao a sensor ka lebelo le tsitsitseng le le phahameng, ho hlahisa lintlha tse ngata esita le ha ho se na lisebelisoa tse molemo. Ho feta moo, mekhoa ena e sebelisa mekhoa e rarahaneng ea ho sebetsana le matšoao a dijithale ho etsa esale pele data ea ho kenya (hangata e lerata). Instead, biology offers alternative solutions for processing noisy sensory data using energy-efficient, asynchronous, event-driven approaches (spikes)2,3. Neuromorphic computing takes inspiration from biological systems to reduce computational costs in terms of energy and memory requirements compared to traditional signal processing methods4,5,6. Recently, innovative general purpose brain-based systems implementing impulse neural networks (TrueNorth7, BrainScaleS8, DYNAP-SE9, Loihi10, Spinnaker11) have been demonstrated. Li-processor tsena li fana ka matla a tlase, litharollo tse tlase tsa latency bakeng sa ho ithuta ka mochini le mohlala oa cortical circuit. Ho sebelisa matla a bona a matla ka botlalo, li-processor tsena tsa neuromorphic li tlameha ho hokahana ka kotloloho le li-sensor tse tsamaisoang ke ketsahalo12,13. Leha ho le joalo, kajeno ho na le lisebelisoa tse 'maloa feela tse amang tse fanang ka data e tsamaisoang ke liketsahalo ka kotloloho. Prominent examples are dynamic visual sensors (DVS) for vision applications such as tracking and motion detection14,15,16,17 the silicon cochlea18 and neuromorphic auditory sensors (NAS)19 for auditory signal processing, olfactory sensors20 and numerous examples21,22 of touch. . li-sensor tsa sebopeho.
Leqepheng lena, re fana ka sistimi e ncha e ntlafalitsoeng e tsamaisoang ke ketsahalo e sebelisoang molemong oa ho etsa lintho. Here, for the first time, we describe an end-to-end system for object localization obtained by connecting a state-of-the-art piezoelectric micromachined ultrasonic transducer (pMUT) with a computational graph based on neuromorphic resistive memory (RRAM). Mehaho ea k'homphieutha ea mohopolong e sebelisang RRAM ke tharollo e tšepisang bakeng sa ho fokotsa tšebeliso ea matla23,24,25,26,27,28,29. Their inherent non-volatility—not requiring active power consumption to store or update information—is a perfect fit with the asynchronous, event-driven nature of neuromorphic computing, resulting in near-no power consumption when the system is idle. Li-piezoelectric micromachined ultrasonic transducers (pMUTs) li theko e tlaase, li-miniaturized silicon-based ultrasonic transducers tse khonang ho sebetsa e le li-transmitters le ba amohelang30,31,32,33,34. Ho sebetsana le mats'oao a amohetsoeng ke li-sensor tse hahelletsoeng ka hare, re ile ra khothatsoa ke moliko oa sephooko neuroanatomy35,36,37. Sephooko se bitsoang Tyto alba se tsebahala ka bokhoni ba sona bo hlollang ba ho tsoma bosiu ka lebaka la mokhoa o sebetsang hantle oa ho utloa. To calculate the location of prey, the barn owl's localization system encodes the time of flight (ToF) when sound waves from prey reach each of the owl's ears or sound receptors. Given the distance between the ears, the difference between the two ToF measurements (Interaural Time Difference, ITD) makes it possible to analytically calculate the azimuth position of the target. Although biological systems are poorly suited to solving algebraic equations, they can solve localization problems very effectively. The barn owl nervous system uses a set of coincidence detector (CD)35 neurons (ie, neurons capable of detecting temporal correlations between spikes that propagate downward to convergent excitatory endings)38,39 organized into computational graphs to solve positioning problems.
Previous research has shown that complementary metal-oxide-semiconductor (CMOS) hardware and RRAM-based neuromorphic hardware inspired by the inferior colliculus (“auditory cortex”) of the barn owl is an efficient method for calculating position using ITD13, 40, 41, 42, 43 , 44, 45, 46. However, the potential of complete neuromorphic systems that link auditory cues to neuromorphic computational graphs has yet to be demonstrated. Bothata bo boholo ke phapang ea tlhaho ea li-circuits tsa analog tsa CMOS, tse amang ho nepahala ha tlhahlobo ea papali. Haufinyane tjena, mekhoa e meng ea lipalo ea likhakanyo tsa ITD47 e bontšitsoe. In this paper, we propose to use the ability of RRAM to change the conductance value in a non-volatile manner to counteract variability in analog circuits. We implemented an experimental system consisting of one pMUT transmitting membrane operating at a frequency of 111.9 kHz, two pMUT receiving membranes (sensors) simulating barn owl ears, and one . We experimentally characterized the pMUT detection system and RRAM-based ITD computational graph to test our localization system and evaluate its angular resolution.
Re bapisa mokhoa oa rona le ts'ebetsong ea dijithale ho microcontroller e etsang mosebetsi o ts'oanang oa sebaka ka ho sebelisa mekhoa e tloaelehileng ea beamforming kapa neuromorphic, hammoho le field programmable gate array (FPGA) bakeng sa khakanyo ea ITD e hlahisitsoeng bukeng. 47. Papiso ena e totobatsa katleho ea matla a tlhōlisano ea mokhoa o reriloeng oa RRAM-based analog neuromorphic system.

sephooko sa moliko se amohela maqhubu a molumo ho tsoa ho sepheo, tabeng ena se tsamaisa phofu. The time of flight (ToF) of the sound wave is different for each ear (unless the prey is directly in front of the owl). The dotted line shows the path that sound waves take to reach the barn owl's ears. Prey can be accurately localized in the horizontal plane based on the length difference between the two acoustic paths and the corresponding interaural time difference (ITD) (left image inspired by ref. 74, copyright 2002, Society for Neuroscience). Sistimi ea rona, transmitter ea pMUT (boputsoa bo lefifi) e hlahisa maqhubu a molumo a tlokomang moo sepheo sa sona. Maqhubu a bonts'itsoeng a ultrasound a amoheloa ke ba amohelang pMUT ba babeli (botala bo khanyang) mme ba sebetsoa ke processor ea neuromorphic (ka ho le letona). b Moetso oa khomphutha oa ITD (Jeffress) o hlalosang kamoo melumo e kenang litsebeng tsa morubisi oa moliko e kenngoeng pele e le li-spikes tse notletsoeng ka har'a nucleus e kholo (NM) ebe ho sebelisoa marang-rang a hlophisitsoeng ka geometrical a li-neurone tsa detector tse tsamaisanang khubung ea lamellar. Ho sebetsa (Netherlands) (ka ho le letšehali). Papiso ea graph ea computational ea neuroITD e kopanyang mela ea ho lieha le li-neurone tsa ho lemoha lintho tse etsahetseng ka tšohanyetso, sistimi ea sephooko sa biosensor e ka etsoa mohlala ho sebelisoa li-circuits tsa neuromorphic tse thehiloeng ho RRAM (ka ho le letona). 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). Joale CD e tla kenyelletsa boemo ba angular ea sepheo.
Piezoelectric micromechanical ultrasonic transducers ke scalable ultrasonic transducers e ka kopanngoang le theknoloji e tsoetseng pele ea CMOS31,32,33,52 'me e na le matla a tlaase a pele a matla le tšebeliso ea matla ho feta li-transducers tse tloaelehileng tsa volumetric53. Mosebetsing oa rona, bophara ba membrane ke 880 µm, 'me maqhubu a resonant a ajoa ka har'a 110-117 kHz (Setšoantšo sa 2a, bona Mekhoa bakeng sa lintlha). Sehlopheng sa lisebelisoa tse leshome tsa tlhahlobo, boleng bo tloaelehileng bo ne bo ka ba 50 (ref. 31). Theknoloji e fihlile khōlong ea indasteri 'me ha e na bioinspired per se. Ho kopanya tlhahisoleseding e tsoang lifiliming tse fapaneng tsa pMUT ke mokhoa o tsebahalang, 'me tlhahisoleseding ea angle e ka fumanoa ho pMUTs ho sebelisoa, mohlala, mekhoa ea ho betla31,54. Leha ho le joalo, ts'ebetso ea mats'oao e hlokahalang ho ntša lintlha tsa angle ha e loketse litekanyo tse tlase tsa matla. Sistimi e reriloeng e kopanya "neuromorphic data preprocessing circuit pMUT le graph ea RRAM-based neuromorphic computing e bululetsoeng ke mohlala oa Jeffress (Figure 2c), e fanang ka tharollo ea hardware e sebetsang hantle le e nang le lisebelisoa. Re entse teko eo ho eona li-sensor tse peli tsa pMUT li neng li behiloe karohano e ka bang 10 cm ho sebelisa melumo e fapaneng ea ToF e amoheloang ke li-membrane tse peli tse amohelang. pMUT e le 'ngoe e sebetsang e le transmitter e lula pakeng tsa ba amohelang. Sepheo e ne e le poleiti ea PVC e bophara ba 12 cm, e leng hole D ka pel'a sesebelisoa sa pMUT (setšoantšo sa 2b). Motho ea amohelang o tlaleha molumo o hlahang ho ntho eo 'me o itšoara ka hohle kamoo ho ka khonehang nakong ea ho feta ha leqhubu la molumo. Pheta teko ka ho fetola boemo ba ntho, bo laoloang ke sebaka sa D le angle θ. E bululetsoe ke sehokelo. 55, re sisinya ts'ebetso ea neuromorphic pele ho ts'ebetso ea matšoao a tala a pMUT ho fetolela maqhubu a bonahatsoang hore e be litlhoro ho kenya graph ea computational ea neuromorphic. ToF e tsamaellanang le amplitude ea tlhōrō e ntšoa ho e 'ngoe le e' ngoe ea litsela tse peli 'me e kenyelelitsoe e le nako e nepahetseng ea litlhōrō ka bomong. Ka feiga. 2c e bonts'a potoloho e hlokahalang ho hokahanya sensor ea pMUT ka graph ea computational e thehiloeng ho RRAM: bakeng sa e 'ngoe le e' ngoe ea ba amohelang pMUT e 'ngoe le e' ngoe, lets'oao le tala le hloekisoa hore le boreleli, le lokisoe, ebe le fetisetsoa ho sehokelo se lutlang ka mokhoa oa ho hlola. the dynamic threshold (Fig. 2d) e etsa ketsahalo ea tlhahiso (spike) le firing (LIF) neuron: nako ea spike ea tlhahiso e kenyelletsa nako ea sefofane e fumanoeng. Boemo ba LIF bo lekantsoe khahlanong le karabelo ea pMUT, kahoo e fokotsa phapano ea pMUT ho tloha sesebelisoa ho ea ho sesebelisoa. Ka mokhoa ona, ho e-na le ho boloka leqhubu lohle la molumo mohopolong le ho le sebetsa hamorao, re mpa re hlahisa tlhōrō e tsamaellanang le ToF ea leqhubu la molumo, e leng ho kenya letsoho ho graphive memory computational graph. Li-spikes li romelloa ka ho toba meleng ea tieho 'me li tsamaisana le li-module tsa ho lemoha lipapali ho li-graph tsa neuromorphic computation. Hobane li romelloa lihekeng tsa li-transistors, ha ho na potoloho e eketsehileng ea amplification e hlokahalang (sheba Supplementary Fig. 4 bakeng sa lintlha tse qaqileng). E le ho lekola ho nepahala ha li-angular tsa sebaka seo ho fanoeng ke pMUT le mokhoa o reriloeng oa ho sebetsana le matšoao, re ile ra lekanya ITD (ke hore, phapang ea nako pakeng tsa liketsahalo tse phahameng tse hlahisoang ke ba amohelang batho ba babeli) ha sebaka le sekhutlo sa ntho e fapana. Tlhahlobo ea ITD e ile ea fetoloa hore e be li-angles (sheba Mekhoa) 'me e reriloe khahlanong le boemo ba ntho: ho se ts'oanehe ho ITD e lekantsoeng ho ile ha eketseha ka sebaka le sekhutlo ho ntho (setšoantšo sa 2e, f). Bothata bo boholo ke peak-to-noise ratio (PNR) karabong ea pMUT. Ntho e ka thōko, e fokotsa pontšo ea acoustic, kahoo e fokotsa PNR (setšoantšo sa 2f, mola o motala). Ho fokotseha ha PNR ho lebisa ho eketseha ha ho se ts'oanehe ha tekanyo ea ITD, e leng se etsang hore ho be le keketseho ea ho nepahala ha libaka (setšoantšo sa 2f, mola o moputsoa). Bakeng sa ntho e ka thōko ho 50 cm ho tloha ho transmitter, ho nepahala ha angular ea tsamaiso ke hoo e ka bang 10 °. Moeli ona o behiloeng ke litšobotsi tsa sensor o ka ntlafatsoa. Ka mohlala, khatello e rometsoeng ke emitter e ka eketsoa, ​​​​ka hona e eketsa motlakase o tsamaisang lesela la pMUT. Tharollo e 'ngoe ea ho matlafatsa letšoao le fetisitsoeng ke ho hokahanya li-transmitters tse ngata 56. Litharollo tsena li tla eketsa sebaka sa ho lemoha ka litšenyehelo tsa litšenyehelo tse eketsehileng tsa matla. Lintlafatso tse ling li ka etsoa lehlakoreng la ho amohela. Lebato la lerata la moamoheli oa pMUT le ka fokotsoa haholo ka ho ntlafatsa khokahano lipakeng tsa pMUT le amplifier ea mohato oa pele, e ntseng e etsoa hajoale ka likhokahanyo tsa terata le likhoele tsa RJ45.
Setšoantšo sa kristale ea pMUT e nang le li-membrane tse tšeletseng tsa 880 µm tse kopantsoeng ka bophahamo ba 1.5 mm. b Setšoantšo sa mokhoa oa ho lekanya. The target is located at azimuth position θ and at distance D. The pMUT transmitter generates a 117.6 kHz signal that bounces off the target and reaches two pMUT receivers with different time-of-flight (ToF). This difference, defined as the inter-aural time difference (ITD), encodes the position of an object and can be estimated by estimating the peak response of the two receiver sensors. c Schematic ea mehato ea pele ho ts'ebetso ea ho fetolela lets'oao le tala la pMUT hore e be tatellano ea spike (ke hore, ho kenya graph ea neuromorphic computation). The pMUT sensors and neuromorphic computational graphs have been fabricated and tested, and the neuromorphic pre-processing is based on software simulation. d Karabelo ea lesela la pMUT ha le fumana lets'oao le ho fetoha sebaka sa spike. e Ho nepahala ha sebaka sa liteko ka mokhoa oa ts'ebetso ea angle ea ntho (Θ) le sebaka (D) ho ea ho sepheo. Mokhoa oa ho ntša oa ITD o hloka bonyane qeto ea angular e ka bang 4°C. f Ho nepahala ha angular (mothapo o moputsoa) le karo-karolelano ea tlhoro-ho-lerata (mola o motala) khahlano le sebaka sa ntho bakeng sa Θ = 0.
Resistive memory e boloka tlhahisoleseding sebakeng se sa feto-fetoheng sa conductive. Molao-motheo oa motheo oa mokhoa ona ke hore ho fetoloa ha thepa boemong ba athomo ho baka phetoho ho conductivity ea eona ea motlakase57. Mona re sebelisa mohopolo oa ho hanyetsa o thehiloeng ho oxide o nang le lesela la 5nm la hafnium dioxide e pakeng tsa titanium e ka holimo le e ka tlaase le li-electrode tsa titanium nitride. Ts'ebetso ea lisebelisoa tsa RRAM e ka fetoloa ka ho sebelisa waveform ea hona joale / ea motlakase e hlahisang kapa e senyang likhoele tse tsamaisang likheo tsa oksijene lipakeng tsa li-electrode. Re ile ra kopanya lisebelisoa tse joalo58 ka mokhoa o tloaelehileng oa 130 nm CMOS ho theha potoloho e hlophisitsoeng e nchafalitsoeng ea neuromorphic e kenyang ts'ebetso ea detector ea iketsahalletseng le potoloho ea ho lieha (Setšoantšo sa 3a). Sebopeho se sa tsitsang le sa analoge sa sesebelisoa, se kopantsoe le tlhaho e tsamaisoang ke ketsahalo ea potoloho ea neuromorphic, e fokotsa tšebeliso ea matla. Potoloho e na le ts'ebetso ea hang-hang / ea ho tima: e sebetsa hang ka mor'a ho buloa, ho lumella matla ho tima ka ho feletseng ha potoloho e sa sebetse. Libaka tse ka sehloohong tsa mohaho oa morero o reriloeng li bontšitsoe setšoantšong. 3b. E na le meaho ea N parallel single-resistor single-transistor (1T1R) e kenyelletsang litekanyo tsa synaptic moo maqhubu a boima a nkuoang, a kenngoa ka har'a synapse e tloaelehileng ea "different pair integrator" (DPI) 59, 'me qetellong e kenngoa ka har'a synapse ka ho kopanya le ho kopanya. dutla. activated (LIF) neuron 60 (sheba Mekhoa bakeng sa lintlha). Lits'oants'o tse kenang li sebelisoa hekeng ea sebopeho sa 1T1R ka mokhoa oa tatellano ea li-voltage pulses ka nako ka tatellano ea makholo a li-nanoseconds. Mehopolo e hanyetsanang e ka behoa sebakeng se phahameng sa conductive (HCS) ka ho sebelisa ts'upiso e ntle ea kantle ho Vtop ha Vbottom e theiloe, 'me e khutlisetsoe boemong bo tlaase ba conductive (LCS) ka ho sebelisa voltage e ntle ho Vbottom ha Vtop e thehiloe. Karolelano ea boleng ba HCS e ka laoloa ka ho fokotsa lenaneo la hona joale (ho lumellana) le SET (ICC) ka motlakase oa heke-mohloli oa letoto la transistor (setšoantšo sa 3c). Mesebetsi ea RRAM potolohong e habeli: e tsamaisa le ho lekanya matla a ho kenya letsoho.
Setshwantsho sa electron microscope (SEM) sa sesebediswa se seputswa sa HfO2 1T1R RRAM se kopantsweng ho theknoloji ya 130 nm CMOS e nang le selector transistors (650 nm wide) ka botala. b Mehaho ea motheo ea schema ea neuromorphic e reriloeng. Li-pulses tsa motlakase oa ho kenya (lihloro) Vin0 le Vin1 li sebelisa Iweight ea hona joale, e lekanang le libaka tsa conduction G0 le G1 ea sebopeho sa 1T1R. Hona joale e kenngoa ka har'a li-synapse tsa DPI 'me e thabisa li-neuron tsa LIF. RRAM G0 le G1 li kentsoe ho HCS le LCS ka ho latellana. c Mosebetsi oa cumulative conductance density bakeng sa sehlopha sa lisebelisoa tsa 16K RRAM e le ts'ebetso ea ICC e bapisang hona joale, e laolang hantle boemo ba conduction. d Litekanyo tsa potoloho ho (a) tse bonts'ang hore G1 (ho LCS) e thibela ka nepo ho kenya letsoho ho tsoa ho Vin1 (e tala), 'me ehlile voltage ea membrane ea neuron e arabela feela ho kenyelletso e putsoa ho tsoa ho Vin0. RRAM e khetholla ka nepo likhokahano tsa potoloho. e Tekanyo ea potoloho ka (b) e bontšang phello ea boleng ba conductance G0 ka lera la Vmem ka mor'a ho sebelisa voltage pulse Vin0. Ha boitšoaro bo ntse bo eketseha, karabelo e matlafala: ka hona, sesebelisoa sa RRAM se sebelisa boima ba khokahano ea I/O. Litekanyo li entsoe potolohong 'me li bonts'a ts'ebetso e habeli ea RRAM, ho tsamaisa le ho lekanya boima ba li-pulse tse kenang.
First, since there are two basic conduction states (HCS and LCS), RRAMs can block or miss input pulses when they are in the LCS or HCS states, respectively. Ka lebaka leo, RRAM e khetholla ka nepo likhokahano tsa potoloho. This is the basis for being able to reconfigure the architecture. To demonstrate this, we will describe a fabricated circuit implementation of the circuit block in Fig. 3b. RRAM e tsamaellanang le G0 e hlophiselitsoe ho HCS, 'me RRAM G1 ea bobeli e kentsoe LCS. Li-pulse tse kenyang li sebelisoa ho Vin0 le Vin1 ka bobeli. 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. Teko e ile ea atleha ha feela sesebediswa sa HCS (G0) se ne se hoketswe ho pulse ya neuron ho tsosa tsitsipano ya lera. Sena se bontšoa ho Setšoantšo sa 3d, moo terene e putsoa e etsang hore motlakase oa membrane o hahe holim'a capacitor ea membrane, athe terene e tala e boloka motlakase oa membrane o sa fetohe.
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. Tekong ea bobeli, sesebelisoa sa G0 se ile sa hlophisoa maemong a fapaneng a HCS, 'me phallo ea ho kenya e sebelisoa ho kenya letsoho ho VIn0. The input pulse draws a current (Iweight) from the device, which is proportional to the conductance and the corresponding potential drop Vtop − Vbot. Hona joale e boima e kenngoa ka har'a li-synapse tsa DPI le li-neuron tsa LIF. Motlakase oa membrane oa li-neuron tse hlahisoang o ngotsoe ho sebelisoa oscilloscope 'me o bontšitsoe setšoantšong sa 3d. Tlhōrō ea "voltage" ea membrane ea neuron karabelong ea "pulse" e le 'ngoe e lekana le ts'ebetso ea mohopolo o hanyetsanang, ho bonts'a hore RRAM e ka sebelisoa e le ntho e ka hlophisehang ea boima ba synaptic. Liteko tsena tse peli tsa selelekela li bonts'a hore sethala se reriloeng sa RRAM-based neuromorphic se khona ho kenya tšebetsong lintlha tsa motheo tsa mochini oa Jeffress oa mantlha, e leng mohala oa ho lieha le potoloho ea tsietsi. The circuit platform is built by stacking successive blocks side by side, such as the blocks in Figure 3b, and connecting their gates to a common input line. We designed, fabricated, and tested a neuromorphic platform consisting of two output neurons receiving two inputs (Fig. 4a). Setšoantšo sa potoloho se bontšoa setšoantšong sa 4b. Matrix a 2 × 2 RRAM e ka holimo e lumella li-pulse tsa ho kenya hore li lebisoe ho li-neuron tse peli tse hlahisoang, ha 2 × 2 matrix e ka tlaase e lumella likhokahano tse tloaelehileng tsa li-neurons tse peli (N0, N1). Re bontša hore sethala sena se ka sebelisoa ka tlhophiso ea mohala oa ho lieha le mesebetsi e 'meli e fapaneng ea ho lemoha lintho tse etsahetseng ka tšohanyetso, joalokaha ho bontšoa ke litekanyo tsa liteko ho Feiga 4c-e.
Circuit diagram formed by two output neurons N0 and N1 receiving two inputs 0 and 1. The top four devices of the array define synaptic connections from input to output, and the bottom four cells define recurrent connections between neurons. The colored RRAMs represent the devices configured in the HCS on the right: the devices in the HCS allow connections and represent weights, while the devices in the LCS block input pulses and disable connections to outputs. b Diagram of circuit (a) with eight RRAM modules highlighted in blue. c Mela ea tieho e thehoa ka ho sebelisa feela matla a DPI synapses le LIF neurons. RRAM e tala e behiloe ho sebetsa holimo ho lekana hore e khone ho baka glitch sephethong ka mor'a ho lieha ho kenya Δt. d Papiso ea moralo ea ho lemoha ha CD e sa tsotelleng tataiso ea matšoao a itšetlehileng ka nako. Output neuron 1, N1, e tuka ho kenya letsoho 0 le 1 ka tieho e kgutshwane. e Direction sensitive CD circuit, potoloho e lemohang ha ho kenya letsoho 1 ho atamela ho kenya 0 mme e fihla ka mor'a ho kenya 0. Phallo ea potoloho e emeloa ke neuron 1 (N1).

Ho feto-fetoha ke mohloli oa ho se phethahale ho mekhoa ea methapo ea methapo63,64,65. Sena se lebisa boitšoarong bo fapaneng ba li-neurone le li-synapses. Mehlala ea bofokoli bo joalo e kenyelletsa 30% (ho kheloha ho tloaelehileng) ho feto-fetoha ha phaello ea ho kenya letsoho, nako e tsitsitseng, le nako ea refractory, ho bolela tse seng kae feela (sheba Mekhoa). Bothata bona bo bonahala le ho feta ha li-circuits tse ngata tsa methapo ea kutlo li hokahantsoe hammoho, joalo ka CD e utloang maikutlo e nang le li-neurone tse peli. Ho sebetsa hantle, nako ea phaello le ho bola ea li-neuron tse peli li lokela ho tšoana ka hohle kamoo ho ka khonehang. Mohlala, phapang e kholo ea phaello ea ho kenya letsoho e ka etsa hore neuron e le 'ngoe e sebetse ho feta tekano ha neuron e 'ngoe e sa arabele hantle. Ka feiga. Setšoantšo sa 5a se bontša hore li-neurone tse khethiloeng ka nako e sa lekanyetsoang li arabela ka tsela e fapaneng ho phallo e tšoanang ea ho kenya. Phapang ena ea methapo e bohlokoa, mohlala, mosebetsing oa li-CD tse utloang tataiso. Lenaneong le bontšitsoeng ho feiga. 5b, c, the input gain of neuron 1 is much higher than that of neuron 0. Thus, neuron 0 requires three input pulses (instead of 1) to reach the threshold, and neuron 1, as expected, needs two input events. Implementing spike time-dependent biomimetic plasticity (STDP) is a possible way to mitigate the impact of imprecise and sluggish neural and synaptic circuits on system performance43. Here we propose to use the plastic behavior of resistive memory as a means of influencing the enhancement of neural input and reducing the effects of variability in neuromorphic circuits. Joalokaha ho bontšitsoe feiga. 4e, maemo a boitšoaro a amanang le RRAM synaptic mass a ile a fetola ka katleho karabo e lumellanang ea neural membrane voltage. Re sebelisa leano le pheta-phetoang la RRAM. For a given input, the conductance values ​​of the synaptic weights are reprogrammed until the target behavior of the circuit is obtained (see Methods).
Litekanyo tsa liteko tsa karabelo ea li-neuron tse robong tse khethiloeng ka mokhoa o sa reroang ho phallo e tšoanang ea ho kenya. The response varies across populations, affecting input gain and time constant. b Experimental measurements of the influence of neurons on the variability of neurons affecting direction-sensitive CD. The two direction-sensitive CD output neurons respond differently to input stimuli due to neuron-to-neuron variability. Neuron 0 e na le phaello e tlase ea ho kenya ho feta neuron 1, kahoo ho nka li-pulse tse tharo tsa ho kenya (ho e-na le 1) ho theha sekhahla sa tlhahiso. Joalokaha ho lebelletsoe, neuron 1 e fihla monyako ka liketsahalo tse peli tsa ho kenya. Haeba input 1 e fihla Δt = 50 µs kamora ho tuka ha neuron 0, CD e lula e khutsitse hobane Δt e kholo ho feta nako e sa fetoheng ea neuron 1 (hoo e ka bang 22 µs). c e fokotsoa ke Δt = 20 µs, e le hore ho kenya 1 tlhōrō ha ho thunngoa ha neuron 1 ho ntse ho le holimo, ho fella ka ho lemoha ka nako e le 'ngoe liketsahalo tse peli tsa ho kenya.

Phello ea ho feto-fetoha ha neuronal ho li-circuits tsa tieho. b Lipotoloho tsa mela e liehang li ka lekanyetsoa ho tieho e kholo ka ho beha linako tsa nako ea li-neuron tsa LIF le li-synapse tsa DPI ho litekanyetso tse kholo. Ho eketsa palo ea ho pheta-pheta ts'ebetso ea ho lekanya ea RRAM ho entse hore ho khonehe ho ntlafatsa haholo ho nepahala ha sepheo sa ho lieha: ho pheta-pheta ha 200 ho fokolitse phoso ka tlase ho 5%. Phetoho e le 'ngoe e tsamaisana le ts'ebetso ea SET/RESET ho sele ea RRAM. Mojule o mong le o mong oa CD oa mofuta oa c Jeffress o ka kengoa ts'ebetsong ho sebelisoa likarolo tsa N parallel CD bakeng sa ho fetoha le maemo ho hoholo mabapi le mefokolo ea sistimi. d Liphetoho tse ling tsa RRAM tsa calibration li eketsa sekhahla sa 'nete se nepahetseng (mola o moputsoa), ha sekhahla se fosahetseng se itšetlehile ka palo ea ho pheta-pheta (mola o motala). Ho beha likarolo tse ling tsa CD ka tsela e tšoanang ho qoba ho fumanoa ka bohata ha li-module tsa CD.
Hona joale re hlahloba ts'ebetso le ts'ebeliso ea matla ea mokhoa oa ho qetela oa ho qetela o kopantsoeng oa ntho ea sebaka se bontšitsoeng ho Setšoantšo sa 2 ho sebelisa litekanyo tsa thepa ea acoustic ea pMUT sensor, CD, le lipotoloho tsa ho lieha ho etsa li-neuromorphic computing graph. Mohlala oa Jeffress (setšoantšo sa 1a). Ha e le graph ea computing ea neuromorphic, palo e kholo ea li-module tsa CD, e ntlafatsa tharollo ea angular, empa hape e phahamisa matla a tsamaiso (setšoantšo sa 7a). Ho sekisetsa ho ka finyelloa ka ho bapisa ho nepahala ha likarolo tsa motho ka mong (li-sensor tsa pMUT, neurons, le li-circuits tsa synaptic) le ho nepahala ha tsamaiso eohle. Qeto ea mohala oa ho lieha e lekanyelitsoe ke linako tsa nako ea li-synapse le li-neuron tse etsisitsoeng, tseo morerong oa rona o fetang 10 µs, o lumellanang le qeto ea angular ea 4 ° (bona Mekhoa). Li-node tse tsoetseng pele haholoanyane tse nang le theknoloji ea CMOS li tla lumella moralo oa li-circuits tsa neural le synaptic tse nang le li-constants tsa nako e tlase, e leng se hlahisang ho nepahala ho hoholo ha likarolo tsa mohala oa ho lieha. Leha ho le joalo, tsamaisong ea rona, ho nepahala ho lekanyelitsoe ke phoso pMUT ka ho lekanya boemo ba angular, ke hore 10 ° (mohala o moputsoa o otlolohileng setšoantšong sa 7a). Re ile ra lokisa palo ea li-module tsa CD ho 40, e lumellanang le qeto ea angular e ka bang 4 °, ke hore, ho nepahala ha angular ea graph computational (mola o khanyang o moputsoa o otlolohileng setšoantšong sa 7a). Boemong ba sistimi, sena se fana ka qeto ea 4 ° le ho nepahala ha 10 ° bakeng sa lintho tse fumanehang 50 cm ka pel'a sistimi ea sensor. Boleng bona bo bapisoa le lits'ebetso tsa "neuromorphic sound localization" tse tlalehiloeng ho ref. 67. Papiso ea tsamaiso e reriloeng le boemo ba boemo bo botle e ka fumanoa ho Letlapa la Tlatsetso 1. Ho eketsa pMUT e eketsehileng, ho eketsa boemo ba pontšo ea acoustic, le ho fokotsa lerata la elektronike ke litsela tse ka khonehang tsa ho ntlafatsa ho nepahala ha libaka. ) e hakanyetsoa ho 9.7. nz. 55. Ho fanoe ka li-CD tse 40 tsa graph ea computational, papiso ea SPICE e hakantse matla ka ts'ebetso e 'ngoe le e' ngoe (ke hore, matla a ho beha ntho) ho 21.6 nJ. Sisteme ea neuromorphic e kengoa tšebetsong feela ha ketsahalo ea ho kenya e fihla, ke hore, ha leqhubu la acoustic le fihla moamoheli ofe kapa ofe oa pMUT mme le feta moeli oa ho lemoha, ho seng joalo le lula le sa sebetse. Sena se qoba tšebeliso ea matla e sa hlokahaleng ha ho se na letšoao la ho kenya. Ha ho nahanoa ka makhetlo a ts'ebetso ea sebaka sa 100 Hz le nako ea ts'ebetso ea 300 µs ka ts'ebetso (tekanyo e phahameng ka ho fetisisa ea ITD), tšebeliso ea matla ea neuromorphic computing graph ke 61.7 nW. Ka neuromorphic pre-processing e sebelisoang ho moamoheli e mong le e mong oa pMUT, tšebeliso ea matla ea sistimi eohle e fihla ho 81.6 nW. Ho utloisisa ts'ebetso e ntle ea matla ea mokhoa o reriloeng oa neuromorphic ha o bapisoa le lisebelisoa tse tloaelehileng, re bapisitse palo ena le matla a hlokahalang ho etsa mosebetsi o ts'oanang ho "microcontroller" ea sejoale-joale ea matla a tlase e sebelisang bokhoni ba neuromorphic kapa bo tloaelehileng ba beamforming68. Mokhoa oa neuromorphic o nka sethala sa analog-to-digital converter (ADC), se lateloa ke sefahla sa li-band-pass le sethala sa ho ntša enfelopo (mokhoa oa Teeger-Kaiser). Qetellong, ho etsoa opereishene ea ho ntša ToF. Re tlohetse lipalo tsa ITD tse ipapisitseng le ToF le phetoho ho boemo bo hakantsoeng ba angular kaha sena se etsahala hang bakeng sa tekanyo ka 'ngoe (bona Mekhoa). Ha re nka lisampole tsa 250 kHz likanaleng ka bobeli (li-receiver tsa pMUT), ts'ebetso ea filthara ea li-band tse 18, ts'ebetso ea ho ntša enfelopo e 3, le ts'ebetso e le 'ngoe ea sampuli ka' ngoe, tšebeliso ea matla e hakanyetsoa ho li-microwatts tse 245. Sena se sebelisa microcontroller's low-power mode69, e bulelang ha li-algorithms li sa sebetse, e fokotsang tšebeliso ea matla ho 10.8 µW. Tšebeliso ea matla ea tharollo ea mats'oao a beamforming e hlahisitsoeng ho referense. 31, e nang le baamoheli ba 5 pMUT le maballo a 11 a ajoang ka mokhoa o ts'oanang sefofaneng sa azimuth [-50 °, +50 °], ke 11.71 mW (sheba karolo ea Mekhoa bakeng sa lintlha). Ntle le moo, re tlaleha ts'ebeliso ea matla ea FPGA47-based Time Difference Encoder (TDE) e hakantsoeng ho 1.5 mW e le sebaka sa mofuta oa Jeffress bakeng sa ho etsa ntho ea lehae. Ho ipapisitsoe le likhakanyo tsena, mokhoa o reriloeng oa neuromorphic o fokotsa ts'ebeliso ea matla ka liodara tse hlano tsa boholo ha o bapisoa le microcontroller e sebelisang mekhoa ea khale ea beamforming bakeng sa ts'ebetso ea sebaka sa ntho. Ho amohela mokhoa oa neuromorphic oa ts'ebetso ea lets'oao ho microcontroller ea khale ho fokotsa tšebeliso ea matla ka liodara tse peli tsa boholo. Ho sebetsa hantle ha sistimi e reriloeng ho ka hlalosoa ka ho kopanngoa ha potoloho ea analoge ea asynchronous resistive-memory e khonang ho etsa lipalo tsa mohopolong le khaello ea phetoho ea analoge ho ea ho dijithale e hlokahalang ho lemoha matšoao.
qeto ea Angular (e putsoa) le tšebeliso ea matla (e tala) ea ts'ebetso ea libaka ho itšetlehile ka palo ea li-module tsa CD. The dark blue horizontal bar represents the angular accuracy of the PMUT and the light blue horizontal bar represents the angular accuracy of the neuromorphic computational graph. b Power consumption of the proposed system and comparison with the two discussed microcontroller implementations and digital implementation of the Time Difference Encoder (TDE)47 FPGA.
Ho fokotsa ts'ebeliso ea matla a sistimi e reriloeng ea sebaka, re ile ra theha, ra theha le ho kenya tšebetsong potoloho e sebetsang hantle, e tsamaisoang ke ketsahalo ea RRAM-based neuromorphic e sebetsanang le tlhaiso-leseling e hlahisoang ke li-sensor tse hahelletsoeng ho bala boemo ba ntho e lebisitsoeng ka nepo. nako. . Le ha mekhoa ea khale ea ts'ebetso e ntse e etsa mohlala oa matšoao a fumanoeng le ho etsa lipalo ho fumana tlhaiso-leseling e sebetsang, tharollo e reriloeng ea neuromorphic e etsa lipalo ka mokhoa o ts'oanang ha tlhaiso-leseling e bohlokoa e fihla, e holisa ts'ebetso ea matla a sistimi ka liodara tse hlano tsa boholo. Ho phaella moo, re totobatsa ho feto-fetoha ha li-circuits tsa neuromorphic tse thehiloeng ho RRAM. Bokhoni ba RRAM ba ho fetola boitšoaro ka mokhoa o sa tsitsang (polasetiki) bo lefella ho fetoha ha tlhaho ha li-circuits tsa synaptic le methapo ea methapo ea DPI. Sena se etsa hore potoloho ena e thehiloeng ho RRAM e feto-fetohe le ho ba matla. Sepheo sa rona ha se ho ntša mesebetsi e rarahaneng kapa lipaterone ho tsoa ho matšoao, empa ke ho beha lintho sebakeng ka nako ea nnete. Sistimi ea rona le eona e ka hatella lets'oao hantle mme qetellong ea le romella mehatong e tsoelang pele ea ts'ebetso ho etsa liqeto tse rarahaneng ha ho hlokahala. Boemong ba lits'ebetso tsa sebaka sa marang-rang, mohato oa rona oa "neuromorphic preprocessing" o ka fana ka leseli mabapi le sebaka sa lintho. Lintlha tsena li ka sebelisoa, mohlala, bakeng sa ho lemoha motsamao kapa ho lemoha boitšisinyo. Re totobatsa bohlokoa ba ho kopanya li-sensor tsa motlakase tse tlase haholo tse kang pMUT le lisebelisoa tsa elektroniki tse tlase haholo. Bakeng sa sena, mekhoa ea neuromorphic e bile ea bohlokoa kaha e re lebisitse ho nts'etsopele ea ts'ebetso e ncha ea potoloho ea mekhoa ea khomphutha e bululetsoeng ke baeloji joalo ka mohlala oa Jeffress. In the context of sensor fusion applications, our system can be combined with several different event-based sensors to obtain more accurate information. Although owls are excellent at finding prey in the dark, they have excellent eyesight and perform a combined auditory and visual search before catching prey70. When a particular auditory neuron fires, the owl receives the information it needs to determine in which direction to start its visual search, thus focusing its attention on a small part of the visual scene. A combination of visual sensors (DVS camera) and a proposed listening sensor (based on pMUT) should be explored for the development of future autonomous agents.
Sensor ea pMUT e ho PCB e nang le li-receiver tse peli tse arohaneng ka 10 cm, 'me transmitter e teng pakeng tsa ba amohelang. Mosebetsing ona, lera ka 'ngoe ke sebopeho sa bimorph se emisitsoeng se nang le likarolo tse peli tsa piezoelectric aluminium nitride (AlN) 800 nm e teteaneng e pakeng tsa likarolo tse tharo tsa molybdenum (Mo) 200 nm e teteaneng 'me e koahetsoe ka lera la 200 nm e teteaneng. karolo e kaholimo ea SiN joalo ka ha e hlalositsoe ho referense. 71. Li-electrode tse ka hare le tse ka ntle li sebelisoa ka tlaase le ka holimo ho likarolo tsa molybdenum, ha electrode e bohareng ea molybdenum e sa tloaeleha ebile e sebelisoa e le fatše, e leng se hlahisang lera le nang le lipara tse 'nè tsa li-electrode.
Mohaho ona o lumella tšebeliso ea "membrane deformation" e tloaelehileng, e hlahisang phetiso e ntlafalitsoeng le ho amohela kutloisiso. PMUT e joalo hangata e bonts'a maikutlo a ho hlasimolla a 700 nm/V joalo ka emitter, e fana ka khatello e holimo ea 270 Pa/V. As a receiver, one pMUT film exhibits a short circuit sensitivity of 15 nA/Pa, which is directly related to the piezoelectric coefficient of AlN. Phapang ea tekheniki ea motlakase sebakeng sa AlN e lebisa phetohong ea maqhubu a resonant, a ka lefelloang ka ho sebelisa leeme la DC ho pMUT. Kutloelo-bohloko ea DC e lekantsoe ho 0.5 kHz/V. Bakeng sa sebopeho sa acoustic, microphone e sebelisoa ka pel'a pMUT.
To measure the echo pulse, we placed a rectangular plate with an area of ​​about 50 cm2 in front of the pMUT to reflect the emitted sound waves. Both the distance between the plates and the angle relative to the pMUT plane are controlled using special holders. A Tectronix CPX400DP voltage source biases three pMUT membranes, tuning the resonant frequency to 111.9 kHz31, while the transmitters are driven by a Tectronix AFG 3102 pulse generator tuned to the resonant frequency (111.9 kHz) and a duty cycle of 0.01. The currents read from the four output ports of each pMUT receiver are converted to voltages using a special differential current and voltage architecture, and the resulting signals are digitized by the Spektrum data acquisition system. Moeli oa ho lemoha o ne o khetholloa ke ho fumanoa ha matšoao a pMUT tlas'a maemo a fapaneng: re ile ra fetisetsa setšoantšo ho ea libakeng tse fapaneng [30, 40, 50, 60, 80, 100] cm mme ra fetola angle ea tšehetso ea pMUT ([0, 20, 40] o ) Setšoantšo sa 2b se bonts'a qeto ea nakoana ea tlhahlobo ea ITD ho itšetlehile ka boemo bo lumellanang ba angular ka likhato.
Sengoliloeng sena se sebelisa lipotoloho tse peli tse fapaneng tsa RRAM tse kantle ho sethala. Ea pele ke lisebelisoa tse ngata tsa 16,384 (16,000) (lisebelisoa tse 128 × 128) ka tlhophiso ea 1T1R e nang le transistor e le 'ngoe le mochine o le mong. Chip ea bobeli ke sethala sa neuromorphic se bontšitsoeng setšoantšong sa 4a. Sele ea RRAM e na le filimi ea 5 nm e tenya ea HfO2 e kentsoeng ka har'a stack ea TiN/HfO2/Ti/TiN. The RRAM stack is integrated into the back-of-line (BEOL) of the standard 130nm CMOS process. RRAM-based neuromorphic circuits present a design challenge for all-analog electronic systems in which RRAM devices coexist with traditional CMOS technology. Haholo-holo, boemo ba ho tsamaisa sesebelisoa sa RRAM bo tlameha ho baloa le ho sebelisoa e le phetoho ea ts'ebetso bakeng sa sistimi. Ho finyella sena, potoloho e ne e entsoe, e entsoe le ho lekoa e balang hona joale ho tloha mochine ha ho amoheloa pulse ea ho kenya 'me e sebelisa hona joale ho lekanya karabo ea synapse ea "dipharol pair integrator" (DPI). Potoloho ena e bontšoa ho Setšoantšo sa 3a, se emelang li-block tsa motheo tsa sethala sa neuromorphic ho Setšoantšo sa 4a. Pulse e kenang e kenya ts'ebetsong heke ea sesebelisoa sa 1T1R, e etsa hore e be hona joale ka RRAM e lekanang le tsamaiso ea mochine G (Iweight = G (Vtop - Vx)). Kenyelletso e kenang ea potoloho ea amplifier (op-amp) e sebetsang e na le Vtop e sa fetoheng ea DC bias voltage. Maikutlo a fosahetseng a op-amp a tla fana ka Vx = Vtop ka ho fana ka tekanyo e lekanang ea hona joale ho tloha ho M1. Iweight ea morao-rao e nkiloeng sesebelisoa e kenngoa ka har'a synapse ea DPI. Matla a matla a hona joale a tla fella ka depolarization e eketsehileng, kahoo tsamaiso ea RRAM e sebelisa litekanyo tsa synaptic ka katleho. Hona joale ea exponential synaptic current e kenngoa ka membrane capacitor ea Leaky Integration and Excitation (LIF) neurons, moo e kopantsoeng e le motlakase. Haeba threshold voltage ea membrane (the switching voltage of the inverter) e hloloa, karolo e hlahisoang ke neuron e ea sebetsa, e hlahisa sekhahla sa tlhahiso. This pulse returns and shunts the neuron's membrane capacitor to ground, causing it to discharge. Potoloho ena e tlatsitsoe ka pulse expander (e sa bonts'itsoeng setšoantšong sa 3a), e bōpang phallo e hlahisoang ke neuron ea LIF ho ea bophara ba sepheo sa sekhahla. Multiplexers are also built into each line, allowing voltage to be applied to the top and bottom electrodes of the RRAM device.
Teko ea motlakase e kenyelletsa ho sekaseka le ho rekota boits'oaro bo matla ba li-circuits tsa analog, hammoho le ho etsa mananeo le ho bala lisebelisoa tsa RRAM. Mehato ka bobeli e hloka lisebelisoa tse khethehileng, tseo kaofela li kopantsoeng le boto ea sensor ka nako e le 'ngoe. Ho fihlella lisebelisoa tsa RRAM lipotolohong tsa neuromorphic ho etsoa ho tsoa lisebelisoa tsa kantle ka multiplexer (MUX). MUX e arola sele ea 1T1R ho tloha sebakeng se seng sa potoloho eo e leng ho eona, e lumellang sesebelisoa hore se baloe le / kapa se hlophisoe. Ho etsa lenaneo le ho bala lisebelisoa tsa RRAM, mochine oa Keithley 4200 SCS o sebelisoa hammoho le microcontroller ea Arduino: ea pele bakeng sa moloko o nepahetseng oa pulse le ho bala hona joale, 'me oa bobeli ke oa ho fihlella ka potlako ho likarolo tsa 1T1R ka mokhoa oa mohopolo. Ts'ebetso ea pele ke ho theha sesebelisoa sa RRAM. Lisele li khethiloe ka bonngoe 'me motlakase o motle o sebelisoa pakeng tsa li-electrode tse ka holimo le tse ka tlaase. Tabeng ena, hona joale ho lekanyelitsoe ho tatellano ea mashome a microamperes ka lebaka la phepelo ea motlakase oa heke e lumellanang le transistor ea khetho. Ka nako eo sele ea RRAM e ka potoloha pakeng tsa boemo bo tlaase ba conductive (LCS) le boemo bo phahameng ba ho khanna (HCS) ho sebelisa ts'ebetso ea RESET le SET, ka ho latellana. Ts'ebetso ea SET e etsoa ka ho sebelisa sekhahla sa motlakase oa kgutlonne ka nako ea 1 μs le tlhōrō ea motlakase oa 2.0-2.5 V ho electrode e ka holimo, le sync pulse ea sebopeho se tšoanang le tlhōrō ea 0.9-1.3 V ho ea ho heke ea selector transistor. Litekanyetso tsena li lumella ho tsamaisa RRAM ka nako ea 20-150 µs. Bakeng sa RESET, bophara ba 1 µs, 3 V peak pulse e sebelisoa ho electrode e ka tlase (bit line) ea sele ha motlakase oa heke o le ka har'a 2.5-3.0 V. Lintho tse kenang le tse hlahisoang ke li-circuits tsa analog ke matšoao a matla. . Bakeng sa ho kenya letsoho, re kentse lijenereithara tse peli tsa HP 8110 le lijenereithara tsa matšoao tsa Tektronix AFG3011. Pulse ea ho kenya e na le bophara ba 1 µs le moeli oa ho nyoloha / oa ho theoha oa 50 ns. Mofuta ona oa pulse o nkoa e le glitch e tloaelehileng ho li-circuits tse thehiloeng ho analog. Ha e le lets'oao la tlhahiso, lets'oao la tlhahiso le tlalehiloe ho sebelisoa Teledyne LeCroy 1 GHz oscilloscope. Lebelo la ho fumana oscilloscope le netefalitsoe hore ha se lebaka le lekanyelitsoeng ho tlhahlobo le ho fumana data ea potoloho.
Using the dynamics of analog electronics to simulate the behavior of neurons and synapses is an elegant and efficient solution to improve computational efficiency. Bothata ba taba ena ea computational underlay ke hore e tla fapana ho ea ka morero. Re ile ra lekanya phapang ea li-neurons le li-circuits tsa synaptic (Supplementary Fig. 2a,b). Har'a liponahatso tsohle tsa ho fetoha, tse amanang le nako e tsitsitseng le phaello ea ho kenya letsoho li na le tšusumetso e kholo ka ho fetisisa boemong ba tsamaiso. The time constant of the LIF neuron and the DPI synapse is determined by an RC circuit, where the value of R is controlled by a bias voltage applied to the gate of the transistor (Vlk for the neuron and Vtau for the synapse), determining the sekhahla sa ho dutla. Input gain is defined as the peak voltage reached by the synaptic and neuronal membrane capacitors stimulated by an input pulse. Phaello ea ho kenya e laoloa ke transistor e 'ngoe ea bias e fetolang motlakase oa hona joale. Papiso ea Monte Carlo e lekantsoeng tšebetsong ea ST Microelectronics' 130nm e ile ea etsoa ho bokella phaello e itseng le lipalo-palo tse sa fetoheng tsa nako. The results are presented in Supplementary Figure 2, where the input gain and time constant are quantified as a function of the bias voltage controlling the leakage rate. Matšoao a matala a lekanya phapang e tloaelehileng ea nako e sa fetoheng ho tloha ho moelelo. 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. Kenyelletso ea ho kenya letsoho (Supplementary Fig. 2e,d) ea phapang ea neuronal le synapse e ne e ka ba 8% le 3%, ka ho latellana. Khaello e joalo e ngotsoe hantle ka har'a lingoliloeng: litekanyo tse fapaneng li entsoe ka bongata ba lichifi tsa DYNAP ho lekola phapang lipakeng tsa li-neurone tsa LIF63. Li-synapse tse ho BrainScale chip ea matšoao a tsoakiloeng a ile a lekanngoa 'me ho se lumellane ha bona ho ile ha hlahlojoa,' me ho ile ha etsoa tlhahiso ea mokhoa oa ho lekanya ho fokotsa phello ea ho fetoha ha maemo a tsamaiso64.
Mosebetsi oa RRAM lipotolohong tsa neuromorphic e habeli: tlhaloso ea meralo (ho kenya litsamaiso ho liphetho) le ts'ebetsong ea boima ba synaptic. Thepa ea ho qetela e ka sebelisoa ho rarolla bothata ba ho fetoha ha li-circuits tsa neuromorphic tsa mohlala. We have developed a simple calibration procedure that involves reprogramming the RRAM device until the circuit being analyzed meets certain requirements. For a given input, the output is monitored and the RRAM is reprogrammed until the target behavior is achieved. A wait time of 5 s was introduced between programming operations to solve the problem of RRAM relaxation resulting in transient conductance fluctuations (Supplementary Information). Boima ba synaptic boa fetoloa kapa bo lekanngoe ho latela litlhoko tsa potoloho ea neuromorphic e etsoang mohlala. The calibration procedure is summarized in additional algorithms [1, 2] that focus on two fundamental features of neuromorphic platforms, delay lines and direction insensitive CD. Bakeng sa potoloho e nang le mohala oa ho lieha, sepheo sa sepheo ke ho fana ka sekhahla se hlahisoang ka tieho Δt. If the actual circuit delay is less than the target value, the synaptic weight of G3 should be reduced (G3 should be reset and then set to a lower matching current Icc). Conversely, if the actual delay is greater than the target value, the conductance of G3 must be increased (G3 must first be reset and then set to a higher Icc value). Ts'ebetso ena e phetoa ho fihlela ho lieha ho hlahisoang ke potoloho ho lumellana le boleng ba sepheo mme ho behoa mamello ho emisa ts'ebetso ea ho lekanya. For orientation-insensitive CDs, two RRAM devices, G1 and G3, are involved in the calibration process. This circuit has two inputs, Vin0 and Vin1, delayed by dt. The circuit should only respond to delays below the matching range [0,dtCD]. If there is no output peak, but the input peak is close, both RRAM devices should be boosted to help the neuron reach the threshold. Conversely, if the circuit responds to a delay that exceeds the target range of dtCD, the conductance must be reduced. Pheta ts'ebetso ho fihlela boitšoaro bo nepahetseng bo fumanoa. Hona joale ho lumellana ho ka fetoloa ka potoloho ea analoge e hahiloeng ho ref. 72.73. Ka potoloho ena e hahelletsoeng, lits'ebetso tse joalo li ka etsoa nako le nako ho lekanya sistimi kapa ho e sebelisa hape bakeng sa ts'ebeliso e 'ngoe.

Lintlha tse tšehetsang liphetho tsa phuputso ena li fumaneha ho tsoa ho sengoli se amehang, FM, ka kopo e utloahalang.
Indiveri, G. & Sandamirskaya, Y. Bohlokoa ba sebaka le nako bakeng sa ts'ebetso ea matšoao ho li-neuromorphic agents: Phephetso ea ho hlahisa matla a tlaase, mahlahana a ikemetseng a sebelisanang le tikoloho. Indiveri, G. & Sandamirskaya, Y. Bohlokoa ba sebaka le nako bakeng sa ts'ebetso ea matšoao ho li-neuromorphic agents: Phephetso ea ho hlahisa matla a tlaase, mahlahana a ikemetseng a sebelisanang le tikoloho.Indiveri G. le Sandamirskaya Y. Bohlokoa ba sebaka le nako bakeng sa ts'ebetso ea matšoao ho li-neuromorphic agents: phephetso ea ho hlahisa li-agent tse ikemetseng tse nang le matla a tlaase a sebelisanang le tikoloho. Indiverri, G. & Sandamirskaya, Y. 空间和时间对于神经形态代理中信号处理 的 Indiveri, G. & Sandamirskaya, Y.Indiveri G. le Sandamirskaya Y. Bohlokoa ba sebaka le nako bakeng sa ts'ebetso ea matšoao ho li-neuromorphic agents: phephetso ea ho hlahisa li-agent tse ikemetseng tse nang le matla a tlaase a sebelisanang le tikoloho.IEEE Signal Processing. Journal 36, 16–28 (2019).
Thorpe, SJ Peak Nako ea ho Fihla: Neural Network Coding Scheme e Sebetsang. ho Eckmiller, R., Hartmann, G. & Hauske, G. (eds). ho Eckmiller, R., Hartmann, G. & Hauske, G. (eds).ho Eckmiller, R., Hartmann, G. le Hauske, G. (eds.).Ho Eckmiller, R., Hartmann, G., le Hauske, G. (eds.). Ts'ebetso e ts'oanang ea lits'ebetso tsa neural le likhomphutha 91-94 (North-Holland Elsevier, 1990).
Levy, WB & Calvert, VG Communication e sebelisa matla a 35 ho feta computation ka cortex ea motho, empa litšenyehelo ka bobeli lia hlokahala ho bolela esale pele nomoro ea synapse. Levy, WB & Calvert, VG Communication e sebelisa matla a 35 ho feta computation ka cortex ea motho, empa litšenyehelo ka bobeli lia hlokahala ho bolela esale pele nomoro ea synapse.Levy, WB le Calvert, Puisano ea WG e sebelisa matla a fetang 35 ho feta computation ka cortex ea motho, empa litšenyehelo ka bobeli lia hlokahala ho bolela esale pele palo ea li-synapse. 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 PuisanoLevy, WB le Calvert, Puisano ea WG e sebelisa matla a makhetlo a 35 ho feta computation ka cortex ea motho, empa litšenyehelo ka bobeli li hloka ho bolela esale pele palo ea li-synapses.tshebetso. Sekolo sa Sechaba sa Saense. saense. US 118, https://doi.org/10.1073/pnas.2008173118 (2021).
Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. Computing neuromorphic e bululetsoeng ke likokoanyana. Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. Computing neuromorphic e bululetsoeng ke likokoanyana.Dalgati, T., Vianello, E., DeSalvo, B. le Casas, J. Computing neuromorphic e bululetsoeng ke likokoanyana.Dalgati T., Vianello E., DeSalvo B. le Casas J. Computing neuromorphic e bululetsoeng ke likokoanyana. Hona joale. Maikutlo. Saense ea likokoanyana. 30, 59-66 (2018).
Roy, K., Jaiswal, A. & Panda, P. Towards spike-based machine intelligence with neuromorphic computing. Roy, K., Jaiswal, A. & Panda, P. Towards spike-based machine intelligence with neuromorphic computing. Roy, K., Jaiswal, A. & Panda, P. Towards Spike-Based Machine Intelligence with Neuromorphic Computing.Roy K, Jaiswal A, le Panda P. Pulse-based intelligence ea maiketsetso e sebelisang neuromorphic computing. Tlhaho 575, 607–617 (2019).
Indiveri, G. & Liu, S.-C. Indiveri, G. & Liu, S.-C.Indiveri, G. le Liu, S.-K. Indiveri, G. & Liu, S.-C. Indiveri, G. & Liu, S.-C.Indiveri, G. le Liu, S.-K.Memori le ts'ebetso ea tlhahisoleseling lits'ebetsong tsa neuromorphic. tshebetso. IEEE 103, 1379–1397 (2015).
Akopyan F. et al. Truenorth: Moralo le sephutheloana sa lisebelisoa bakeng sa 65 mW 1 million neuron programmable synaptic chip. Lits'ebetso tsa IEEE. Moralo oa khomphutha oa litsamaiso tse kopaneng tsa potoloho. 34, 1537–1557 (2015).
Schemmel, J. le ba bang. Live demo: scaled down version of the BrainScaleS neuromorphic system at plate scale. 2012 IEEE International Symposium on Circuits and Systems (ISCAS), (IEEE ed.) 702–702 (2012).
Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. A scalable multicore architecture e nang le mekhoa e mengata ea memori bakeng sa li-processor tsa neuromorphic asynchronous (DYNAPs). Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. A scalable multicore architecture e nang le mekhoa e mengata ea memori bakeng sa li-processor tsa neuromorphic asynchronous (DYNAPs).Moradi S., Qiao N., Stefanini F. le Indiviri G. A scalable multicore architecture le heterogeneous memory structures for dynamic neuromorphic asynchronous processors (DYNAP). Moradi, S, Qiao, N. Moradi, S., Qiao, N., Stefanini, F. & Indiveri, G. Mofuta oa meaho e atolositsoeng ea li-multi-core, e nang le sebopeho se ikhethileng sa mohopolo bakeng sa ts'ebetso ea neural e matla (DYNAP).Moradi S., Qiao N., Stefanini F. le Indiviri G. A scalable multicore architecture le heterogeneous memory structures for dynamic neuromorphic asynchronous processors (DYNAP).IEEE Transactions ka Biomedical science. tsamaiso ea motlakase. 12, 106–122 (2018).
Davis, M. et al. Loihi: A neuromorphic multi-core processor with embedded learning. IEEE Micro 38, 82–99 (2018).
Furber, SB, Galluppi, F., Temple, S. & Plana, LA Morero oa SpiNNaker. Furber, SB, Galluppi, F., Temple, S. & Plana, LA Morero oa SpiNNaker.Ferber SB, Galluppi F., Temple S. and Plana LA SpiNNaker project. tshebetso. IEEE 102, 652-665 (2014).
Liu, S.-K. & Delbruck, T. Neuromorphic sensory systems. & Delbruck, T. Neuromorphic sensory systems.le Delbrück T. Neuromorphic sensory systems. & Delbruck, T. 神经形态感觉系统。 & Delbruck, T.le Delbrück T. Neuromorphic sensory system.Hona joale. Maikutlo. Neurobiology. 20, 288-295 (2010).
Chope, T. et al. Neuromorphic sensory integration bakeng sa ho kopanngoa ha mohloli oa molumo le ho qoba ho thulana. Ka 2019 Sebokeng sa IEEE mabapi le Lipotoloho le Litsamaiso tsa Biomedical (BioCAS), (IEEE Ed.) 1–4 (2019).
Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G. A spike-based neuromorphic architecture ea stereo vision. Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G. A spike-based neuromorphic architecture ea stereo vision.Risi N, Aymar A, Donati E, Solinas S, le Indiveri G. A 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. & Indiveri, G.Risi N, Aimar A, Donati E, Solinas S, le Indiveri G. Spike-based neuromorphic architecture bakeng sa pono ea stereo.ka pele. Neurorobotics 14, 93 (2020).
Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. A spiking neural network model of 3Dperception bakeng sa liketsahalo tse thehiloeng ho neuromorphic stereo vision systems. Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. A spiking neural network model of 3Dperception bakeng sa liketsahalo tse thehiloeng ho neuromorphic stereo vision systems.Oswald, M., Ieng, S.-H., Benosman, R., le Indiveri, G. A 3D Pulsed Neural Network Perception Model bakeng sa Ketsahalo e thehiloeng ho 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。 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., le Indiveri, G. Spiked 3Dperception Neural Network Model bakeng sa Ketsahalo e thehiloeng ho Neuromorphic Stereo Vision System.saense. Tlaleho 7, 1-11 (2017).
Dalgaty, T. et al. Ho lemoha motsamao oa mantlha o susumetsoang ke likokoanyana ho kenyelletsa mohopolo o hananang le o mong le marang-rang a phatlohileng a methapo ea kutlo. Sistimi ea biohybrid ea biohybrid. 10928, 115-128 (2018).
D'Angelo, G. et al. Ho lemoha motsamao o ipapisitseng le ketsahalo o sebelisa likhoutu tsa nakoana. ka pele. Neurology. La 14, 451 (2020).