Iindaba

Real-world data processing applications require compact, low-latency, low-power computing systems. Ngobuchule bekhompyuter obuqhutywa ngumnyhadala, i-metal-oxide-semiconductor hybrid memristive neuromorphic architectures ibonelela ngesiseko esifanelekileyo sehardware kwimisebenzi enjalo. Ukubonisa amandla apheleleyo eenkqubo ezinjalo, sicebisa kwaye sibonise ngovavanyo isisombululo esibanzi sokusetyenzwa kwenzwa yosetyenziso lwento yendawo yokwenyani. 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 ezi ziphumo zovavanyo ukulungelelanisa ukulinganisa kwinqanaba lenkqubo. Ezi zifaniso zisetyenziselwa ukuvavanya isisombululo se-angular kunye nokusebenza kakuhle kwamandla kwemodeli yendawo yendawo. Iziphumo zibonisa ukuba indlela yethu inokuba zii-odolo ezininzi zobukhulu obusebenza ngokufanelekileyo kunee-microcontrollers ezenza umsebenzi ofanayo.
Singena kwixesha le-computing yonke indawo apho inani lezixhobo kunye neenkqubo ezisetyenzisiweyo zikhula ngokukhawuleza ukusinceda kubomi bethu bemihla ngemihla. Ezi nkqubo zilindeleke ukuba ziqhube ngokuqhubekayo, zidla amandla amancinci njengoko kunokwenzeka ngelixa zifunda ukutolika idatha abayiqokelelayo kwii-sensors ezininzi ngexesha langempela kwaye zivelise imveliso yokubini njengesiphumo sokuhlela okanye ukuqaphela imisebenzi. Elinye lawona manyathelo abalulekileyo afunekayo ukufezekisa le njongo kukukhupha ulwazi oluluncedo noluhlangeneyo olusuka kwingxolo kwaye lusoloko lungaphelelanga kwidatha yeemvakalelo. Conventional engineering approaches typically sample sensor signals at a constant and high rate, generating large amounts of data even in the absence of useful inputs. In addition, these methods use complex digital signal processing techniques to pre-process the (often noisy) input data. Endaweni yoko, ibhayoloji ibonelela ngezisombululo ezizezinye zokusetyenzwa kwedatha yeemvakalelo ezinomsindo kusetyenziswa amandla asebenzayo, asynchronous, iindlela eziqhutywa kumsitho (iispikes)2,3. I-computing ye-Neuromorphic ithatha ukuphefumlelwa kwiinkqubo zebhayoloji ukunciphisa iindleko zokubala ngokweemfuno zamandla kunye nememori xa kuthelekiswa neendlela zendabuko zokucwangcisa iimpawu4,5,6. Kutshanje, iinkqubo ezintsha ezisekelwe kwingqondo ezisekelwe kwingqondo ezisebenzisa i-impulse neural network (TrueNorth7, BrainScaleS8, DYNAP-SE9, Loihi10, Spinnaker11) zibonakalisiwe. These processors provide low power, low latency solutions for machine learning and cortical circuit modeling. To fully exploit their energy efficiency, these neuromorphic processors must be directly connected to event-driven sensors12,13. However, today there are only a few touch devices that directly provide event-driven data. Imizekelo ebalaseleyo zizinzwa eziguquguqukayo ezibonwayo (i-DVS) kwizicelo zombono ezifana nokulandelwa kunye nokubonwa kwentshukumo14,15,16,17 i-silicon cochlea18 kunye ne-neuromorphic auditory sensors (NAS)19 yokwenziwa kwesignali yokuva, i-olfactory sensors20 kunye nemizekelo emininzi21,22 yokuchukumisa. . abenzi boluvo bokuthungwa.
Kweli phepha, sinikezela ngenkqubo esanda kuphuhliswa yokwenziwa kwe-auditory process esetyenziswa kwindawo yento. 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). In-memory computing architectures using RRAM are a promising solution for reducing power consumption23,24,25,26,27,28,29. 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. Piezoelectric micromachined ultrasonic transducers (pMUTs) are inexpensive, miniaturized silicon-based ultrasonic transducers capable of acting as transmitters and receivers30,31,32,33,34. To process the signals received by the built-in sensors, we drew inspiration from barn owl neuroanatomy35,36,37. I-barn owl iTyto alba yaziwa ngobuchule bayo obumangalisayo bokuzingela ebusuku ngenxa yenkqubo esebenzayo yokuvalela indawo. 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. Nangona iinkqubo zebhayoloji zingakufanelanga kakuhle ukusombulula iequation zealjebra, zinokusombulula iingxaki zokwenziwa kwalapha ekhaya ngokuyimpumelelo. 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.

We compare our method with a digital implementation on a microcontroller performing the same localization task using conventional beamforming or neuromorphic methods, as well as a field programmable gate array (FPGA) for ITD estimation proposed in the reference. 47. Olu thelekiso lugxininisa amandla okhuphiswano lwenkqubo ecetywayo ye-analog neuromorphic system ye-RRAM.

the barn owl receives sound waves from a target, in this case moving prey. The time of flight (ToF) of the sound wave is different for each ear (unless the prey is directly in front of the owl). Umgca onamachaphaza ubonisa indlela amaza ahamba ngayo ukuze afikelele ezindlebeni zesikhova. 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). In our system, the pMUT transmitter (dark blue) generates sound waves that bounce off the target. Reflected ultrasound waves are received by two pMUT receivers (light green) and processed by the neuromorphic processor (right). b An ITD (Jeffress) computational model describing how sounds entering the barn owl's ears are first encoded as phase-locked spikes in the large nucleus (NM) and then using a geometrically arranged grid of matched detector neurons in the lamellar nucleus. Processing (Netherlands) (left). Umzobo wegrafu yekhompyutha ye-neuroITD edibanisa imigca yokulibaziseka kunye ne-neurons ye-concidence detector, inkqubo ye-owl biosensor inokulinganiswa kusetyenziswa i-RRAM-based neuromorphic circuits (ekunene). c I-Schematic yeyona ndlela iphambili yeJeffress, ngenxa yomahluko kwi-ToF, iindlebe ezimbini zifumana isivuseleli sesandi ngamaxesha ahlukeneyo kwaye zithumela ii-axon ukusuka kuzo zombini iziphelo ukuya kwisixhobo. 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 iya kuthi emva koko ifake iikhowudi kwindawo ekujoliswe kuyo ye-angular.

umfanekiso wekristale ye-pMUT eneembrane ezithandathu ezingama-880 µm ezidityaniswe kwipitch eyi-1.5 mm. b Diagram of the measuring setup. 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 of pre-processing steps for converting the raw pMUT signal into spike sequences (ie input to the neuromorphic computation graph). The pMUT sensors and neuromorphic computational graphs have been fabricated and tested, and the neuromorphic pre-processing is based on software simulation. d Response of the pMUT membrane upon receipt of a signal and its transformation into a spike domain. e Ukuchaneka kovavanyo lwe-angular njengomsebenzi we-engile yento (Θ) kunye nomgama (D) ukuya kwinto ekujoliswe kuyo. Indlela yokutsalwa kwe-ITD ifuna ubuncinci besisombululo se-angular malunga ne-4°C. f Ukuchaneka kwe-angular (umgca oblowu) kunye nomlinganiselo ohambelana nencopho-kwingxolo (umgca oluhlaza) ngokuchasene nomgama wento Θ = 0.
Imemori exhathisayo igcina ulwazi kwimo yokuqhuba engaguquguqukiyo. Umgaqo osisiseko wendlela kukuba ukuguqulwa kwezinto kumgangatho we-athomu kubangela utshintsho kwi-conductivity yayo yombane57. Here we use an oxide-based resistive memory consisting of a 5nm layer of hafnium dioxide sandwiched between top and bottom titanium and titanium nitride electrodes. The conductivity of RRAM devices can be changed by applying a current/voltage waveform that creates or breaks conductive filaments of oxygen vacancies between the electrodes. We co-integrated such devices58 into a standard 130 nm CMOS process to create a fabricated reconfigurable neuromorphic circuit implementing a coincidence detector and a delay line circuit (Fig. 3a). The non-volatile and analog nature of the device, combined with the event-driven nature of the neuromorphic circuit, minimizes power consumption. The circuit has an instant on/off function: it operates immediately after being turned on, allowing the power to be turned off completely when the circuit is idle. Iibhloko eziphambili zokwakha zesikimu esicetywayo ziboniswe kumkhiwane. 3b. Iqukethe i-N parallel single-resistor single-transistor (1T1R) izakhiwo ezifakela iintsimbi ze-synaptic apho imilinganiselo yokulinganisa ithathwa khona, ifakwe kwi-synapse eqhelekileyo ye-different pair integrator (DPI)59, kwaye ekugqibeleni ifakwe kwi-synapse kunye nokudibanisa kunye leakage. activated (LIF) neuron 60 (see Methods for details). Ukunyuka kwegalelo kusetyenziswe kwisango lesakhiwo se-1T1R ngendlela yokulandelelana kwee-voltage pulses kunye nobude bomyalelo wamakhulu e-nanoseconds. Imemori echasayo inokufakwa kwi-high conductive state (HCS) ngokusebenzisa ireferensi eqinisekileyo yangaphandle kwi-Vtop xa i-Vbottom igxilwe, kwaye iphinde isethelwe kwi-low conductive state (LCS) ngokusebenzisa i-voltage efanelekileyo kwi-Vbottom xa i-Vtop isekelwe. Ixabiso eliqhelekileyo le-HCS linokulawulwa ngokunciphisa inkqubo yangoku (ukuthotyelwa) kwe-SET (ICC) ngesango lomthombo wombane we-series transistor (Fig. 3c). The functions of RRAM in the circuit are twofold: they direct and weight the input pulses.
Ukuskena i-electron microscope (SEM) umfanekiso wesixhobo esiluhlaza se-HfO2 1T1R RRAM esidityaniswe kwi-130 nm iteknoloji ye-CMOS ene-transistors yokukhetha (i-650 nm ububanzi) eluhlaza. b Iibhloko zokwakha ezisisiseko ze-schema ye-neuromorphic ecetywayo. I-voltage input pulses (i-peaks) i-Vin0 kunye ne-Vin1 idla i-Iweight yangoku, ehambelana ne-conductive states G0 kunye ne-G1 yesakhiwo se-1T1R. Le yangoku ifakwe kwi-synapses ye-DPI kwaye ivuyisa i-neurons ye-LIF. I-RRAM G0 kunye ne-G1 zifakwe kwi-HCS kunye ne-LCS ngokulandelelanayo. c Umsebenzi we-cumulative conductance density yeqela lezixhobo ze-16K RRAM njengomsebenzi we-ICC ehambelanayo yangoku, elawula ngokufanelekileyo umgangatho wokuqhuba. d Imilinganiselo yeSekethe ku-(a) ebonisa ukuba i-G1 (kwi-LCS) ithintela ngokufanelekileyo igalelo elivela kwi-Vin1 (eluhlaza), kwaye ngokwenene i-voltage ye-neuron ye-membrane ye-output iphendula kuphela kwigalelo elizuba elivela kwi-Vin0. RRAM effectively determines the connections in the circuit. e Umlinganiselo wesekethe kwi (b) ebonisa umphumo wexabiso le-conductance G0 kwi-membrane voltage Vmem emva kokufaka i-voltage pulse Vin0. Ukuziphatha okungaphezulu, kokukhona impendulo yomelele: ke, isixhobo se-RRAM sisebenzisa ubungakanani boqhagamshelwano lwe-I/O. Imilinganiselo yenziwe kwisekethe kwaye ibonise umsebenzi ombini we-RRAM, umzila kunye nokulinganisa kwee-pulses zokufaka.
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. Ngenxa yoko, i-RRAM imisela ngokufanelekileyo ukudibanisa kwisekethe. Esi sisiseko sokukwazi ukuphinda uqwalasele ulwakhiwo. Ukubonisa oku, siya kuchaza ukuphunyezwa kwesekethe eyenziweyo kwibhloko yesiphaluka kwi-Fig. 3b. I-RRAM ehambelana ne-G0 ifakwe kwi-HCS, kwaye i-RRAM G1 yesibini ifakwe kwi-LCS. Ii-pulses zokufaka zisetyenziswa kuzo zombini i-Vin0 kunye ne-Vin1. Iziphumo zeendlela ezimbini zokulandelelana kwee-pulses zokufaka zahlalutywa kwi-neurons ye-output ngokuqokelela i-voltage ye-neuron membrane kunye nomqondiso wokuphuma usebenzisa i-oscilloscope. Uvavanyo lwaba yimpumelelo xa kuphela isixhobo se-HCS (G0) sasiqhagamshelwe kwi-pulse ye-neuron ukuvuselela ukuxinezeleka kwenwebu. Oku kuboniswa kuMzobo we-3d, apho uloliwe we-blue pulse ubangela ukuba i-voltage ye-membrane yakheke kwi-membrane capacitor, ngelixa i-green pulse train igcina i-voltage ye-membrane ingatshintshi.
Umsebenzi wesibini obalulekileyo we-RRAM kukuphunyezwa kobunzima bokudibanisa. Ukusebenzisa uhlengahlengiso lwe-analog conductance ye-RRAM, udibaniso lwe-I/O lunokulinganiswa ngokufanelekileyo. Kuvavanyo lwesibini, isixhobo se-G0 sacwangciswa kumanqanaba ahlukeneyo e-HCS, kwaye i-pulse yokufaka isetyenziswe kwigalelo le-VIn0. I-pulse yegalelo idonsa i-current (Iweight) ukusuka kwisixhobo, ehambelana nokuqhuba kunye nokuhambelana okunokuthi kuwiswe i-Vtop - Vbot. Le yangoku enesisindo emva koko ifakwe kwi-synapses ye-DPI kunye ne-LIF output neurons. I-voltage ye-membrane ye-neuron ephumayo irekhodwe ngokusebenzisa i-oscilloscope kwaye iboniswe kwi-Fig 3d. Incopho yombane we-membrane ye-neuron ekuphenduleni kwi-pulse yegalelo enye ihambelana nokuqhuba imemori echasayo, ebonisa ukuba i-RRAM ingasetyenziswa njengento ecwangcisiweyo yobunzima be-synaptic. Ezi mvavanyo zimbini zokuqala zibonisa ukuba iqonga elicetywayo le-RRAM-based neuromorphic liyakwazi ukuphumeza izinto ezisisiseko ze-Jeffress mechanism, oko kukuthi umgca wokulibaziseka kunye nesekethe ye-coincidence detector. Iqonga lesekethe lakhiwe ngokubeka iibhloko ezilandelelanayo ngokulandelelana, njengeebhloko kuMzobo 3b, kunye nokudibanisa amasango abo kumgca wokufaka oqhelekileyo. Siye sayila, sayila, saza savavanya iqonga le-neuromorphic eliquka ii-neurons ezimbini eziphumayo ezifumana amagalelo amabini (umzobo 4a). The circuit diagram is shown in Figure 4b. I-2 × 2 i-matrix ye-RRAM ephezulu ivumela i-pulses ye-input ukuba iqondiswe kwii-neurons ezimbini eziphumayo, ngelixa i-2 × i-2 matrix ephantsi ivumela ukuxhamla ngokuphindaphindiweyo kwee-neurons ezimbini (N0, N1). Sibonisa ukuba eli qonga lingasetyenziselwa ukucwangciswa komgca wokulibaziseka kunye nemisebenzi emibini eyahlukeneyo ye-coincidence detector, njengoko kuboniswe ngemilinganiselo yovavanyo kwi-Fig.4c-e.


Variability is a source of imperfection in modeled neuromorphic systems63,64,65. This leads to heterogeneous behavior of neurons and synapses. Imizekelo yezo zinto zingalunganga zibandakanya i-30% (intsingiselo yokutenxa okusemgangathweni) ukuguquguquka kwinzuzo yegalelo, ixesha elingaguqukiyo, kunye nexesha lochaso, ukukhankanya nje ezimbalwa (bona IiNdlela). Le ngxaki ibonakala ngakumbi xa iisekethe ezininzi ze-neural ziqhagamshelwe kunye, njenge-CD ye-orientation-sensitive CD equka ii-neurons ezimbini. Ukusebenza ngokufanelekileyo, inzuzo kunye nexesha lokubola kwee-neurons ezimbini kufuneka zifane ngokusemandleni. Umzekelo, umahluko omkhulu kwingeniso yegalelo unokubangela ukuba i-neuron enye iqhube ngokugqithisileyo kwi-pulse yegalelo ngelixa enye i-neuron ingaphenduli. Kwikhiwane. Umzobo we-5a ubonisa ukuba i-neurons ekhethiweyo ekhethiweyo iphendula ngokungafaniyo kwi-pulse yegalelo elifanayo. Oku kuguquguquka kwe-neural kufanelekileyo, umzekelo, kumsebenzi wee-CDs ze-direction-sensitive CD. In the scheme shown in fig. 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. Apha siphakamisa ukusebenzisa ukuziphatha kweplastiki yememori echasayo njengendlela yokuphembelela ukuphuculwa kwegalelo le-neural kunye nokunciphisa imiphumo yokuguquguquka kwiisekethe ze-neuromorphic. As shown in fig. I-4e, amanqanaba okuqhuba ahambelana ne-RRAM yobunzima be-synaptic aguqule ngokufanelekileyo impendulo yombane we-neural membrane ehambelanayo. We use an iterative RRAM programming strategy. Ngegalelo elinikiweyo, amaxabiso okuqhuba obunzima be-synaptic aphinda acwangciswe kude kufumaneke indlela ekujoliswe kuyo yesekethe (jonga iindlela).
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 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 has a lower input gain than neuron 1, so it takes three input pulses (instead of 1) to create an output spike. Njengoko kulindelekile, i-neuron 1 ifikelela kumda kunye neziganeko ezimbini zokufaka. Ukuba igalelo 1 lifika Δt = 50 µs emva kwemililo ye-neuron 0, i-CD ihlala ithule kuba i-Δt inkulu kunexesha elingaguqukiyo le-neuron 1 (malunga ne-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.

Isiphumo sokuguquguquka kwe-neuronal kwiisekethe zomgca wokulibaziseka. b Delay line circuits can be scaled to large delays by setting the time constants of the corresponding LIF neurons and DPI synapses to large values. Increasing the number of iterations of the RRAM calibration procedure made it possible to significantly improve the accuracy of the target delay: 200 iterations reduced the error to less than 5%. One iteration corresponds to a SET/RESET operation on an RRAM cell. Each CD module in the c Jeffress model can be implemented using N parallel CD elements for greater flexibility with respect to system failures. d More RRAM calibration iterations increase the true positive rate (blue line), while the false positive rate is independent of the number of iterations (green line). Placing more CD elements in parallel avoids false detection of CD module matches.

isisombululo se-Angular (eluhlaza okwesibhakabhaka) kunye nokusetyenziswa kwamandla (okuluhlaza) kokusebenza kwendawo ngokuxhomekeke kwinani leemodyuli zeCD. Ibha ethe tyaba emnyama emnyama imele ukuchaneka kwe-angular ye-PMUT kunye nokukhanya okuluhlaza okwesibhakabhaka okuthe tyaba okuthe tye kubonisa ukuchaneka kwe-angular yegrafu ye-neuromorphic computational. b Ukusetyenziswa kwamandla kwenkqubo ecetywayo kunye nokuthelekisa ezimbini ezixoxwe ngokusetyenziswa kwe-microcontroller kunye nokuphunyezwa kwedijithali kwe-Time Difference Encoder (TDE)47 FPGA.


Olu lwakhiwo luvumela ukusetyenziswa kwe-deformation ye-membrane eqhelekileyo, okubangela ukuhanjiswa okuphuculweyo kunye nokufumana uvakalelo. I-pMUT enjalo idla ngokubonisa uvakalelo lwe-excitation lwe-700 nm/V njenge-emitter, ibonelela ngoxinzelelo lomphezulu we-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. The technical variability of the voltage in the AlN layer leads to a change in the resonant frequency, which can be compensated by applying a DC bias to the pMUT. DC sensitivity was measured at 0.5 kHz/V. Ukwenza uphawu lwe-acoustic, i-microphone isetyenziswa phambi kwe-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. Zombini umgama phakathi kwamacwecwe kunye ne-angle enxulumene nenqwelomoya ye-pMUT ilawulwa kusetyenziswa abanini abakhethekileyo. 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. Imisinga efundwayo ukusuka kumazibuko amane emveliso yomamkeli ngamnye we-pMUT aguqulelwa kumandla ombane kusetyenziswa umahluko okhethekileyo wangoku kunye nolwakhiwo lombane, kwaye iimpawu ezisisiphumo zifakwa kwidijithali yinkqubo yokufunyanwa kwedatha ye-Spektrum. Umda wokufumanisa ubonakaliswe ngokufunyanwa kwesignali ye-pMUT phantsi kweemeko ezahlukeneyo: sihambise isibonisi kwimigama eyahlukeneyo [30, 40, 50, 60, 80, 100] cm kwaye sitshintshe i-angle yokuxhasa i-pMUT ([0, 20, 40] o ) Umzobo we-2b ubonisa isisombululo sokufumanisa i-ITD yesikhashana ngokuxhomekeke kwindawo ehambelanayo ye-angular kwiidigri.
Eli nqaku lisebenzisa iisekethe ezimbini ezahlukeneyo zeRRAM ngaphandle kweshelufu. Eyokuqala luhlu lwezixhobo ze-16,384 (16,000) (izixhobo ze-128 × 128) kwi-1T1R uqwalaselo kunye ne-transistor enye kunye ne-resistor enye. I-chip yesibini yiplatifomu ye-neuromorphic eboniswe kwi-Fig. 4a. Iseli ye-RRAM iqulethe ifilimu ye-5 nm engqindilili ye-HfO2 efakwe kwi-TiN/HfO2/Ti/TiN stack. I-stack ye-RRAM idibaniswe kwi-back-of-line (BEOL) yenkqubo eqhelekileyo ye-130nm ye-CMOS. Iisekethe ze-neuromorphic ezisekwe kwi-RRAM zibonisa umngeni woyilo kuzo zonke iinkqubo zombane ze-analog apho izixhobo ze-RRAM zihlala kunye netekhnoloji ye-CMOS yemveli. Ngokukodwa, imeko yokuqhuba yesixhobo se-RRAM kufuneka ifundwe kwaye isetyenziswe njengenguquko yokusebenza kwesixokelelwano. Ukuza kuthi ga ngoku, isekethe yenzelwe, yenziwe kwaye ivavanywe efunda ikhoyo ngoku kwisixhobo xa i-pulse yegalelo ifunyenwe kwaye isebenzise le yangoku ukulinganisa impendulo ye-different pair integrator (DPI) synapse. Esi sekethe siboniswa kuMfanekiso 3a, omele iibhloko ezisisiseko zesakhiwo se-neuromorphic kwi-Figure 4a. I-pulse yegalelo ivula isango lesixhobo se-1T1R, ibangela i-current ngokusebenzisa i-RRAM ngokulinganayo kwi-conductance yesixhobo G (Isisindo = G (Vtop - Vx)). Igalelo eliguqukayo lesekethe yeamplifier (op-amp) ine-DC bias voltage Vtop engaguqukiyo. Impendulo engalunganga ye-op-amp iya kubonelela ngeVx = Vtop ngokubonelela ngokulinganayo ngoku kwi-M1. I-Iweight yangoku efunyenwe kwisixhobo ifakwe kwi-synapse ye-DPI. A stronger current will result in more depolarization, so RRAM conductance effectively implements synaptic weights. Lo mbane we-synaptic wangoku ujojowe nge-membrane capacitor ye-Leaky Integration kunye ne-Excitation (LIF) neurons, apho idibaniswe njenge-voltage. Ukuba i-voltage ye-threshold ye-membrane (i-voltage yokutshintsha ye-inverter) iyoyiswa, inxalenye ephumayo ye-neuron ivuliwe, ivelisa i-spike ephumayo. This pulse returns and shunts the neuron's membrane capacitor to ground, causing it to discharge. Le sekethi ke yongezwa nge-expander ye-pulse (engaboniswanga kwi-Fig. 3a), eyenza i-pulse ephumayo ye-neuron ye-LIF ukuya kububanzi be-pulse ekujoliswe kuyo. Multiplexers are also built into each line, allowing voltage to be applied to the top and bottom electrodes of the RRAM device.




Idatha exhasa iziphumo zolu phononongo iyafumaneka kumbhali ochaphazelekayo, iFM, ngesicelo esinengqiqo.
I-Indiveri, G. & Sandamirskaya, Y. Ukubaluleka kwendawo kunye nexesha lokwenziwa kwezibonakaliso kwii-agent ze-neuromorphic: Umngeni wokuphuhlisa amandla aphantsi, ama-agent azimeleyo asebenzisana nokusingqongileyo. I-Indiveri, G. & Sandamirskaya, Y. Ukubaluleka kwendawo kunye nexesha lokwenziwa kwezibonakaliso kwii-agent ze-neuromorphic: Umngeni wokuphuhlisa amandla aphantsi, ama-agent azimeleyo asebenzisana nokusingqongileyo.I-Indiveri G. kunye ne-Sandamirskaya Y. Ukubaluleka kwendawo kunye nexesha lokucubungula umqondiso kwii-agent ze-neuromorphic: umngeni wokuphuhlisa i-agent ezizimeleyo ezizimeleyo ezisebenzisana nokusingqongileyo. I-Indieri, G. & Sandandirkyaya, 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. Indiveri, G. & Sandamirskaya, Y.I-Indiveri G. kunye ne-Sandamirskaya Y. Ukubaluleka kwendawo kunye nexesha lokucubungula umqondiso kwii-agent ze-neuromorphic: umngeni wokuphuhlisa i-agent ezizimeleyo ezizimeleyo ezisebenzisana nokusingqongileyo.IEEE Ukwenziwa Kwemiqondiso. Journal 36, 16–28 (2019).
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.).E-Eckmiller, R., Hartmann, G., kunye noHauske, G. (eds.). Ukulungiswa okufanayo kwiinkqubo ze-neural kunye neekhomputha 91-94 (North-Holland Elsevier, 1990).
I-Levy, i-WB kunye ne-Calvert, i-VG yoNxibelelwano idla amaxesha angama-35 ngaphezu kwamandla ngaphezu kokubala kwi-cortex yomntu, kodwa zombini iindleko zifunekayo ukuqikelela inombolo ye-synapse. I-Levy, i-WB kunye ne-Calvert, i-VG yoNxibelelwano idla amaxesha angama-35 ngaphezu kwamandla ngaphezu kokubala kwi-cortex yomntu, kodwa zombini iindleko zifunekayo ukuqikelela inombolo ye-synapse.I-Levy, i-WB kunye neCalvert, i-WG yoNxibelelwano idla amaxesha angama-35 ngaphezu kwamandla ngaphezu kokubala kwi-cortex yomntu, kodwa zombini iindleko zifunekayo ukuqikelela inani le-synapses. Levy, WB & Calvert, VG Unxibelelwano 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 UnxibelelwanoI-Levy, i-WB kunye neCalvert, i-WG yoNxibelelwano idla amaxesha angama-35 ngaphezu kwamandla ngaphezu kokubala kwi-cortex yomntu, kodwa zombini iindleko zifuna ukuqikelela inani le-synapses.inkqubo. IZiko leSizwe leNzululwazi. inzululwazi. I-US 118, https://doi.org/10.1073/pnas.2008173118 (2021).
Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. I-insect-inspired neuromorphic computing. Dalgaty, T., Vianello, E., De Salvo, B. & Casas, J. I-insect-inspired neuromorphic computing.Dalgati, T., Vianello, E., DeSalvo, B. kunye neCasas, J. I-insect-inspired neuromorphic computing.Dalgati T., Vianello E., DeSalvo B. and Casas J. Insect-inspired neuromorphic computing. Current. Uluvo. Inzululwazi yezinambuzane. 30, 59-66 (2018).

I-Indiveri, G. & Liu, S.-C. I-Indiveri, G. & Liu, S.-C.I-Indiveri, uG. kunye noLiu, S.-K. I-Indiveri, G. & Liu, S.-C. I-Indiveri, G. & Liu, S.-C.I-Indiveri, uG. kunye noLiu, S.-K.Imemori kunye nokulungiswa kolwazi kwiinkqubo ze-neuromorphic. inkqubo. IEEE 103, 1379-1397 (2015).
Akopyan F. et al. I-Truenorth: Uyilo kunye nezixhobo zokusebenza ze-65 mW 1 yesigidi se-neuron ecwangcisiweyo ye-synaptic chip. Iintengiselwano zeIEEE. Uyilo lwekhompyuter lweenkqubo ezidibeneyo zeesekethe. 34, 1537-1557 (2015).
Schemmel, J. et al. Idemo ephilayo: inguqulelo eyehlisiwe ye-BrainScaleS neuromorphic system kwisikali sepleyiti. I-2012 IEEE International Symposium yeeSekethe kunye neeNkqubo (ISCAS), (i-IEEE ed.) 702–702 (2012).
IEEE Transactions on Biomedical science. inkqubo yombane. 12, 106-122 (2018).
UDavis, 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 Iprojekthi ye-SpiNNaker. Furber, SB, Galluppi, F., Temple, S. & Plana, LA Iprojekthi ye-SpiNNaker.Ferber SB, Galluppi F., Temple S. kunye neprojekthi yePlana LA SpiNNaker.Ferber SB, Galluppi F., Temple S. kunye neprojekthi yePlana LA SpiNNaker. inkqubo. IEEE 102, 652-665 (2014).
U-Liu, S.-K. & Delbruck, T. Neuromorphic iinkqubo zoluvo. & Delbruck, T. Neuromorphic iinkqubo zoluvo.kunye neDelbrück T. Neuromorphic sensory systems. & Delbruck, T. 神经形态感觉系统。 kunye noDelbruck, T.kunye neDelbrück T. Neuromorphic sensory system.Okwangoku. Uluvo. I-Neurobiology. 20, 288-295 (2010).
Chope, T. et al. Ukuhlanganiswa kwe-Neuromorphic sensory kumthombo odibeneyo womthombo wesandi kunye nokuphepha ukungqubana. Ngo-2019 kwiNkomfa ye-IEEE kwiiSekethe zeBiomedical kunye neNkqubo (i-BioCAS), (i-IEEE Ed.) 1-4 (2019).
Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G. I-spike-based neuromorphic architecture yombono westereo. Risi, N., Aimar, A., Donati, E., Solinas, S. & Indiveri, G. I-spike-based neuromorphic architecture yombono westereo.Risi N, Aymar A, Donati E, Solinas S, kunye 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. & Indiveri, G.Risi N, Aimar A, Donati E, Solinas S, kunye ne-Indiveri G. Spike-based neuromorphic architecture yombono we-stereo.ngaphambili. Neurorobotics 14, 93 (2020).
Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. Imodeli yenethiwekhi ye-spiking ye-3Dperception yeenkqubo zombono ze-neuromorphic stereo. Osswald, M., Ieng, S.-H., Benosman, R. & Indiveri, G. Imodeli yenethiwekhi ye-spiking ye-3Dperception yeenkqubo zombono ze-neuromorphic stereo.Oswald, M., Ieng, S.-H., Benosman, R., kunye ne-Indiveri, G. A 3D Pulsed Neural Network Perception Model yeSiganeko-Based Neuromorphic Vision 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 脉冲神经网络模。 Oswald, M., Ieng, S.-H., Benosman, R., kunye ne-Indiveri, G. Spiked 3Dperception Neural Network Model yeSiganeko-Based Neuromorphic Stereo Vision System.inzululwazi. Ingxelo 7, 1-11 (2017).
Dalgaty, T. et al. Ubhaqo lwentshukumo olusisiseko oluphefumlelwe zisinambuzane lubandakanya inkumbulo exhathisayo kunye nothungelwano lwe-neural olugqabhukileyo. Inkqubo yeBionic biohybrid. 10928, 115–128 (2018).
D'Angelo, G. et al. Ubhaqo lwentshukumo olusekwe kwisiganeko kusetyenziswa ikhowudi yokwahlula okwexeshana. ngaphambili. I-Neurology. 14, 451 (2020).