Teburin Abubuwan Ciki
Matsalar Fasaha
65 nm
Fasahar Sarrafawa
Matsakaicin Ƙarfafawa
1000×
Idan aka kwatanta da Lokacin Halitta
Tallafin Neuron
Hadadden
Dendrites marasa Layi
1. Gabatarwa
Tsarin BrainScaleS (BSS) yana wakiltar ci gaba mai muhimmanci a cikin lissafin kwakwalwar na'ura, yana haɗa aiwatar da samfuran jiki na analog na neurons da synapses tare da cibiyoyin sarrafa dijital. Tsarin BrainScaleS-2 na ƙarni na biyu, wanda aka haɓaka a matsayin wani ɓangare na Aikin Ƙwaƙwalwar Dan Adam na Turai, yana nuna gagarumin ci gaba akan wanda ya gabace shi ta hanyar amfani da fasahar 65 nm da haɗa ƙwararrun raka'o'in sarrafa sassauci na dijital.
2. Bayyani Gabaɗaya na Tsarin BrainScaleS
2.1 Cibiyar Jijiyoyi ta Analog
Cibiyar ta analog tana aiwatar da samfuran jiki na lokaci mai ci gaba na neurons da synapses, tana ba da ingantaccen kwaikwayon cibiyoyin sadarwar jijiyoyi na halitta. Tsarin yana aiki tare da ƙayyadaddun lokaci da yawa ƙanana fiye da tsarin halitta, yana ba da damar saurin kwaikwayon motsin jijiyoyi.
2.2 Sarrafa Sassauci na Dijital
Wani muhimmin ƙirƙira a cikin BSS-2 shine haɗa na'urar sarrafa sassauci ta dijital—mai sarrafa microprocess na musamman wanda aka ƙera musamman don ayyukan koyo a cikin ingantattun tsarin kwakwalwar na'ura na analog. Wannan na'urar tana kula da canje-canjen tsari da sigogi waɗanda ke faruwa akan sassauƙan sassauƙan lokaci idan aka kwatanta da motsin jijiyoyi na analog.
2.3 Ɗabarun Tsarin Kan Guntu (SoC)
Tsarin yana fasalta Tsarin Kan Guntu (SoC) mai kama da kwakwalwa wanda ya ƙunshi cibiyoyin CPU na dijital da yawa tare da keɓaɓɓun raka'o'in vector waɗanda aka haɗa ta hanyar cibiyar sadarwa akan guntu. Wannan ƙirar tana ba da fifikon bayanan abubuwan da suka faru yayin kiyaye sararin samaniya na gama gari na neurons da CPUs.
3. Aiwarar da Fasaha
3.1 HICANN-X ASIC
HICANN-X Application Specific Integrated Circuit yana wakiltar sabon farkon ganuwar tsarin BSS-2. An gina shi cikin fasahar 65 nm, yana ba da damar haɗa hadadden sarrafa dijital tare da da'irorin jijiyoyi na analog.
3.2 Samfuran Neuron da Synapse
Tsarin yana tallafawa samfuran neuron masu sarƙaƙƙiya waɗanda suka haɗa da kwaikwayon tashar ion mai shiryawa da halayen ɓangarorin juna. Wannan yana ba da damar yin samfuri na dendrites marasa layi, matakan aiki masu yaɗawa, NMDA, da yuwuwar Calcium plateau. Ana iya siffanta motsin membrane ta:
$C_m \\frac{dV_m}{dt} = -g_L(V_m - E_L) - \\sum_i g_i(t)(V_m - E_i) + I_{ext}$
3.3 Tsarin Daidaitawa
Akwatin kayan aikin software na al'ada yana sauƙaƙe hadaddun kwaikwayon Monte-Carlo da aka daidaita, yana magance ƙalubalen bambance-bambancen tsari a cikin da'irorin analog. Wannan daidaitawa tana da mahimmanci don nasarar horo da aiki mai dogaro.
4. Sakamakon Gwaji
Tsarin BrainScaleS-2 yana nuna gagarumin ci gaba akan ƙarni na farko. Haɗin sarrafa sassauci na dijital yana ba da damar ƙarin ƙa'idodin koyo masu sassauci fiye da STDP na asali. Na'urar ƙarfafawa ta analog kuma tana tallafawan ninka vector-matrix, yana barin duka fassarar cibiyoyin sadarwa masu zurfi da koyo na gida tare da neurons masu ƙyalli a cikin tushe ɗaya.
Hoto na 1: Abubuwan Tsarin BrainScaleS
Zanen tsarin yana nuna haɗin ma'aunin wafer, BSS-1 ASIC, ƙirar neuron na BSS-2, da misalan alamomin ƙarfin lantarki na membrane waɗanda ke nuna ikon tsarin don kwaikwayon hadaddun motsin jijiyoyi.
Hoto na 2: Tsarin SoC mai kama da kwakwalwa
Tsarin SoC yana kwatanta fale-falen sarrafawa da yawa tare da raka'o'in vector da cibiyoyin analog waɗanda aka haɗa ta hanyar haɗin kai masu faɗi da cibiyar sadarwa akan guntu, yana fasalta fale-falen ayyuka na musamman don sarrafa ƙwaƙwalwar ajiya da SERDES I/O.
5. Aiwarar da Lambar Tsarin (Code)
Tsarin yana amfani da PyNN, harshen siffanta cibiyar sadarwar jijiyoyi marar ƙwaƙwalwar kwaikwayo, yana ba da mu'amalar software ɗaya. A ƙasa akwai misalin sauƙaƙe na saitin neuron:
# Misalin lambar PyNN don BrainScaleS-2
import pyNN.brainscales as bss
# Saita sigogin neuron
neuron_parameters = {
'tau_m': 10.0, # lokacin daidaitawar membrane
'cm': 1.0, # ƙarfin ɗaukar membrane
'v_rest': -70.0, # yuwuwar hutawa
'v_thresh': -55.0, # yuwuwar kofa
'tau_syn_E': 5.0, # lokacin daidaitawar synapse mai tayarwa
'tau_syn_I': 5.0 # lokacin daidaitawar synapse mai hanawa
}
# Ƙirƙiri yawan neuron
population = bss.Population(100, bss.IF_cond_exp, neuron_parameters)
# Saita ƙa'idar sassauci
stdp_model = bss.STDPMechanism(
timing_dependence=bss.SpikePairRule(),
weight_dependence=bss.AdditiveWeightDependence()
)6. Ayyuka na Gaba
Tsarin BrainScaleS-2 yana buɗe sabbin damammaki don ayyukan lissafin kwakwalwar na'ura. Haɗin saurin kwaikwayon analog tare da shirye-shiryen dijital ya sa ya dace da tsarin AI na ainihin lokaci, binciken lissafi mai kwaikwayon kwakwalwa, da ayyukan AI na gefe masu ƙarancin wuta. Ci gaba na gaba na iya mayar da hankali kan sikelin zuwa manyan cibiyoyin sadarwar jijiyoyi, inganta ingantaccen amfani da makamashi, da haɓaka shirye-shiryen ƙa'idodin koyo.
Bincike na Asali
Tsarin BrainScaleS-2 yana wakiltar hanya mai sarƙaƙƙiya ga lissafin kwakwalwar na'ura wanda ke haɗa tazara tsakanin yiwuwar ilimin halitta da ingantaccen lissafi. Ta hanyar haɗa samfuran jiki na analog tare da shirye-shiryen dijital, yana magance muhimman ƙalubale a cikin ƙirar kayan aikin kwakwalwar na'ura. Matsakaicin ƙarfafawar tsarin 1000× idan aka kwatanta da ma'auni na lokacin halitta yana ba da damar ayyukan bincike masu amfani waɗanda in ba haka ba za su buƙaci tsayin lokacin kwaikwayo maras amfani.
Idan aka kwatanta da sauran hanyoyin kwakwalwar na'ura kamar TrueNorth na IBM da Loihi na Intel, BrainScaleS-2 yana ba da fa'idodi na musamman a cikin gaskiyar ilimin halitta ta hanyar aiwatar da sauti. Yayin da tsarin dijital kamar Loihi ke ba da mafi girman shirye-shirye, hanyar analog na BrainScaleS-2 tana ba da mafi kyawun ingantaccen amfani da makamashi don wasu azuzuwan na lissafin jijiyoyi. Wannan ya yi daidai da abubuwan da aka lura a cikin binciken kwakwalwar na'ura na baya-bayan nan, inda hanyoyin haɗin gwiwar analog-dijital ke samun karbuwa saboda daidaitattun halayen aikin su.
Haɗa ƙwararren mai sarrafa sassauci na dijital yana magance babban iyaka na tsarin analog kawai: wahalar aiwatar da hadaddun ƙa'idodin koyo, masu shirya shirye-shirye. Wannan ƙirƙira tana ba BrainScaleS-2 damar tallafawa ba kawai STDP mai ƙarfi ba har ma da ƙarin hanyoyin koyo masu sarƙaƙƙiya, yana mai da shi mafi dacewa don bincike cikin sassauci na jijiyoyi da algorithms koyo.
Tallafin tsarin don duka cibiyoyin sadarwar jijiyoyi masu ƙyalli da fassarar zurfin koyo ta hanyar ninka vector-matrix yana nuna hanya mai ma'ana ga yanayin AI na yanzu. Wannan ikon biyu yana ba masu bincike damar bincika lissafi mai kwaikwayon kwakwalwa yayin kiyaye daidaito da manyan hanyoyin zurfin koyo. Tsarin daidaitawa don sarrafa bambance-bambancen tsarin analog yana nuna ingantaccen injiniyanci wanda ke yarda kuma yana magance ƙalubalen aiki na lissafin kwakwalwar na'ura na analog.
Idan aka duba gaba, tsare-tsare irin na BrainScaleS-2 na iya taka muhimmiyar rawa wajen haɓaka ƙarin tsarin AI masu ingantaccen amfani da makamashi, musamman ma don ayyukan lissafi na gefe inda ƙayyadaddun wutar lantarki ke da mahimmanci. Ci gaba da saka hannun jari na Aikin ƙwaƙwalwar Dan Adam na Turai a cikin wannan fasahar yana jaddada yuwuwar mahimmancinsa ga duka binciken kimiyyar kwakwalwa da ayyukan AI masu amfani.
7. Nassoshi
- Schemmel, J., et al. "A Wafer-Scale Neuromorphic Hardware System for Large-Scale Neural Modeling." ISCAS 2010.
- Indiveri, G., et al. "Neuromorphic silicon neuron circuits." Frontiers in Neuroscience, 2011.
- Davies, M., et al. "Loihi: A Neuromorphic Manycore Processor with On-Chip Learning." IEEE Micro, 2018.
- Merolla, P. A., et al. "A million spiking-neuron integrated circuit with a scalable communication network and interface." Science, 2014.
- Pei, J., et al. "Towards artificial general intelligence with hybrid Tianjic chip architecture." Nature, 2019.
- European Human Brain Project. "Neuromorphic Computing Platform." https://www.humanbrainproject.eu
- IEEE Spectrum. "The Quest for Artificial Intelligence that Mimics the Brain." https://spectrum.ieee.org