Ƙarfafa AI a Bakin Gaci: Ƙwaƙƙwaran Cibiyar Sadarwa ta Binarized da ke dogara da Memristor tare da Kwakwalwan Lantarki na Kusa da Ƙwaƙwalwa

Teburin Abubuwan Ciki

1. Gabatarwa

Binciken ya gabatar da wata hanya mai ban mamaki ga AI a bakin gaci ta hanyar haɗa cibiyoyin sadarwa na binarized na tushen memristor tare da ƙananan ƙwayoyin hasken rana. Wannan haɗin yana magance babbar ƙalubale ta samar da wutar lantarki ga tsarin AI a cikin yanayin da ba a samun ingantattun hanyoyin samar da wutar lantarki ba. Tsarin ya nuna juriya mai ban mamaki ga sauye-sauyen wutar lantarki, yana ci gaba da aiki ko da a ƙarƙashin ƙananan yanayin haske daidai da suns 0.08.

2. Tsarin Fasaha

2.1 Ƙirar Tsarin Memristor

Daidaitawar ta ƙunshi tsararrun memristor guda huɗu, kowanne yana ɗauke da memristor 8,192, wanda ya kai memristor 32,768 gabaɗaya. An tsara kowane tsari a cikin tsari mai giciye wanda aka inganta don kwakwalwan lantarki na dijital na kusa da ƙwaƙwalwa. An ƙera memristors ɗin ta amfani da tsarin samarwa na haɗin CMOS/memristor, wanda ke ba da damar haɗawa mai yawa yayin kiyaye dacewar samarwa da daidaitattun hanyoyin semiconductor.

2.2 Kwakwalwan Lantarki na Dijital na Kusa da Ƙwaƙwalwa

Ba kamar hanyoyin lissafi na analog a cikin ƙwaƙwalwa na al'ada ba, wannan tsarin yana amfani da cikakken tsarin dijital tare na'urar fahimtar ma'ana a cikin na'urar ji da memristors da aka tsara a madadin. Wannan ƙira tana kawar da buƙatar jujjuyawar analog-zuwa-dijital da rikitattun da'irori na kewaye, tana rage amfani da wutar lantarki sosai da kuma inganta juriya ga bambance-bambancen wutar lantarki.

2.3 Tsarin Gudanar da Wutar Lantarki

Tsarin ya haɗa ƙaramin ƙwayar hasken rana mai faɗin bandeji wanda aka inganta musamman don aikace-aikacen cikin gida. An ƙera da'irar gudanar da wutar lantarki don ɗaukar rashin kwanciyar hankali na masu tara makamashi, yana barin cibiyar sadarwa ta canzawa cikin sauƙi tsakanin ingantattun hanyoyin lissafi da kusan dangane da wutar lantarki da ake da ita.

3. Sakamakon Gwaji

3.1 Aiki a Ƙarƙashin Hasken da ya Bambanta

A ƙarƙashin yanayin haske mai ƙarfi, da'irar ta cimma aikin fahimta kwatankwacin wadatar wutar lantarki na benchen ɗakin gwaji, tare da daidaiton rarrabuwa daidai da aiwatar da software. Yayin da haske ya ragu zuwa suns 0.08, tsarin yana ci gaba da aiki tare da raguwar daidaito kaɗan kawai na kashi 8-12 a cikin ma'aunin da aka gwada.

3.2 Daidaito da Amfani da Wutar Lantarki

Binciken ya nuna cewa hotunan da ba a rarraba su da kyau a ƙarƙashin ƙananan wutar lantarki, galibi lamura ne masu wuyar rarrabuwa waɗanda ke ƙalubalantar ko da tsare-tsare masu ƙarfi. Wannan siffa ta raguwa cikin kyau tana sa tsarin ya dace musamman don aikace-aikacen da ake yarda da kura-kurai lokaci-lokaci don musanya tsawon lokacin aiki.

4. Aiwar da Fasaha

4.1 Tushen Lissafi

Cibiyar sadarwa ta binarized tana amfani da ma'auni na binary da kunnawa, tana rage rikitaccen lissafi sosai. Yaɗawar gaba ana iya wakilta shi kamar haka:

$$a^{(l)} = sign(W^{(l)} a^{(l-1)} + b^{(l)})$$

inda $W^{(l)}$ ke wakiltar ma'auni na binary, $a^{(l)}$ kunnawa ne na binary, kuma aikin alama yana fitar da ±1. Giciye na memristor yana yin ninkin matrix $W^{(l)} a^{(l-1)}$ yadda ya kamata ta amfani da lissafi na tushen juriya.

4.2 Aiwar da Lambar

class BinarizedNeuralNetwork:
    def __init__(self, memristor_arrays):
        self.arrays = memristor_arrays
        self.lisa_units = []  # Raka'o'in Fahimtar Ma'ana a cikin Na'urar Ji
        
    def forward_pass(self, input_data):
        # Binarize input
        binary_input = np.sign(input_data)
        
        # Process through memristor arrays
        for i, array in enumerate(self.arrays):
            # Digital near-memory computation
            output = array.compute(binary_input)
            # LISA processing
            output = self.lisa_units[i].process(output)
            binary_input = np.sign(output)
            
        return output
    
    def adaptive_power_mode(self, available_power):
        if available_power < self.power_threshold:
            return "approximate"
        else:
            return "precise"

5. Ayyuka na Guba

Fasahar tana ba da damar yawan aikace-aikace a cikin sa ido na lafiya, amincin masana'antu, da hankalin muhalli. Takamaiman amfani sun haɗa da:

• Na'urorin lura da lafiya masu sarrafa kansu don ci gaba da sa ido akan marasa lafiya
• Masu hankali na hankali don gyaran tsinkaya a cikin saitunan masana'antu
• Tsare-tsaren sa ido na muhalli a wurare masu nisa
• Tsare-tsaren tsaro masu kunna koyaushe tare da iyawar AI da aka saka

Ci gaba na gaba zai iya mayar da hankali kan sanya fasahar zuwa manyan cibiyoyin sadarwa, haɗa hanyoyin tara makamashi da yawa, da haɓaka gine-ginen musamman don takamaiman yankuna na aikace-aikace.

6. Nassoshi

1. Jebali, F. da sauransu. "Ƙarfafa AI a Bakin Gaci: Ƙwaƙƙwaran Cibiyar Sadarwa ta Binarized da ke dogara da Memristor." arXiv:2305.12875 (2023)
2. Hubara, I. da sauransu. "Cibiyoyin Sadarwar Jijiyoyi na Binary." Ci gaba a cikin Tsarin Sarrafa Bayanai na Jijiyoyi (2016)
3. Wong, H. S. P. da sauransu. "RRAM na Metal-oxide." Proceedings of the IEEE (2012)
4. Esser, S. K. da sauransu. "Cibiyoyin sadarwa masu haɗaka don sauri, kwakwalwan lantarki mai ƙarfi." Proceedings of the National Academy of Sciences (2016)
5. Yang, J. J. da sauransu. "Na'urori masu ƙwaƙwalwa don lissafi." Fasahar Nanotechnology (2013)

7. Bincike Mai mahimmanci