Kwamfutocin Neuromorphic don Binciken Hotuna na Tushen Abun Ciki

Teburin Abubuwan Ciki

1. Gabatarwa

Kwamfutocin Neuromorphic suna wakiltar sauyin tsari daga tsarin gine-ginen von Neumann na al'ada ta hanyar kwaikwayon ayyukan jijiyoyi na kwakwalwa ta hanyar hanyoyin sadarwar jijiyoyi masu ƙyalli (SNNs). Wannan binciken yana binciko aikace-aikacen guntuwar Neuromorphic ta Intel Loihi don binciken hotuna na tushen abun ciki (CBIR), yana nuna gagarumin ci gaba a cikin ingancin makamashi yayin kiyaye daidaito mai gasa idan aka kwatanta da na'urori na al'ada.

2. Hanyoyi

2.1 Canjin ANN zuwa SNN

Hanyar ta ƙunshi canza horar da hanyoyin sadarwar jijiyoyi na wucin gadi (ANNs) zuwa hanyoyin sadarwar jijiyoyi masu ƙyalli ta amfani da ɓoyayyen ƙima. Tsarin canjin yana kiyaye iyawar aikin hanyar sadarwa yayin daidaitawa ga yanayin kayan aikin neuromorphic na taron abubuwan da suka faru.

2.2 Tura Loihi

Guntun Loihi na Intel yana aiwatar da SNN tare da kayan aiki na musamman don lissafin jijiyoyi masu ƙyalli. Tsarin turawa ya ƙunshi taswirar SNN da aka canza zuwa neurocores na Loihi da daidaita ka'idojin sadarwar ƙyalli.

3. Aiwar Fasaha

3.1 Tsarin Lissafi

Samfurin neuron mai ƙyalli yana bin ƙwaƙƙwaran haɗa kai da harbi (LIF):

$\tau_m \frac{dV}{dt} = -[V(t) - V_{rest}] + R_m I(t)$

inda $\tau_m$ shine lokacin lokacin membrane, $V(t)$ shine yuwuwar membrane, $V_{rest}$ shine yuwuwar hutawa, $R_m$ shine juriyar membrane, kuma $I(t)$ shine igiyar shigarwa.

3.2 Tsarin Hanyar Sadarwa

Ginin SNN da aka aiwatar ya ƙunshi yadudduka masu jujjuyawa tare da yadudduka masu cikakken haɗin kai. An horar da hanyar sadarwa akan bayanan Fashion-MNIST kuma an daidaita shi don cire fasali a cikin bututun binciken hoto.

4. Sakamakon Gwaji

4.1 Ma'aunin Aiki

Tsarin ya sami daidaitaccen daidaiton dawo da daidai da hanyoyin tushen CNN na al'ada yayin rage yawan amfani da wutar lantarki sosai. Abubuwan da aka samu daga alamu na ƙyalli na lokaci sun tabbatar da yin tasiri don binciken maƙwabta mafi kusa a cikin sararin fasalin gani.

4.2 Binciken Ingancin Makamashi

Binciken kwatancen ya nuna mafita ta neuromorphic ta fi ingancin makamashi sau 2.5 fiye da CPU na ARM Cortex-A72 kuma sau 12.5 mafi inganci fiye da GPU na NVIDIA T4 don ayyukan ƙaddamarwa ba tare da batching ba.

5. Aiwar Lamba

A ƙasa akwai sauƙaƙan lambar ƙarya don bututun binciken hoto na tushen SNN:

# SNN Image Retrieval Pipeline
class SNNImageRetrieval:
    def __init__(self):
        self.snn_model = load_snn_model()
        self.embedding_db = None
    
    def generate_embeddings(self, images):
        """Generate embeddings from spike patterns"""
        embeddings = []
        for img in images:
            spikes = self.snn_model.forward(img)
            embedding = self.extract_spike_features(spikes)
            embeddings.append(embedding)
        return embeddings
    
    def query_image(self, query_img, k=5):
        """Find k nearest neighbors for query image"""
        query_embedding = self.generate_embeddings([query_img])[0]
        distances = cosine_distance(query_embedding, self.embedding_db)
        nearest_indices = np.argsort(distances)[:k]
        return nearest_indices

6. Aikace-aikacen Gaba

Kwamfutocin Neuromorphic suna nuna alƙawari ga aikace-aikacen AI na gefe, binciken bidiyo na ainihin lokaci, da tsarin da aka saka masu ƙarancin wutar lantarki. Hanyoyin bincike na gaba sun haɗa da:

• Haɗa kai tare da gine-ginen transformer don dawo da yanayi daban-daban
• Haɓaka iyawar koyo kan layi don bayanai masu ƙarfi
• Aikace-aikace a cikin tsarin cin gashin kai waɗanda ke buƙatar sarrafa gani na ainihin lokaci
• Haɗuwa tare da algorithms masu wahahin quantum don haɓaka aiki

7. Bincike na Asali

Wannan binciken yana wakiltar babban mataki mai muhimmanci a cikin aikace-aikacen kwamfutocin Neuromorphic don ayyukan hangen nesa na kwamfuta. Ingantaccen ingancin makamashi na 2.5-12.5x da aka nuna akan na'urori na al'ada ya yi daidai da babban yanayin ƙwarewar kayan aikin AI, kama da juyin halitta da aka gani a cikin TPUs na Google da IPUs na Graphcore. Nasarar Loihi a cikin ayyukan dawo da hoto yana nuna cewa gine-ginen neuromorphic na iya zama mai dacewa da tsarin von Neumann na yanzu, musamman don aikace-aikacen lissafi na gefe inda ƙayyadaddun wutar lantarki ke da muhimmanci.

Hanyar canza ANNs da aka riga aka horar da su zuwa SNNs, kamar yadda aka nuna a cikin wannan aikin, tana bin ƙayyadaddun hanyoyin aiki a fagen. Duk da haka, sabon abu yana cikin aiwatar da wannan fasaha musamman ga binciken hoto na tushen abun ciki, aikin da yawanci yana buƙatar albarkatun lissafi masu yawa. Matsakaicin matakan daidaito yayin samun raguwar makamashi mai mahimmanci yana tabbatar da ingancin aikin mafita ta neuromorphic don aikace-aikacen duniya na gaske.

Idan aka kwatanta da sauran ƙirar lissafi masu tasowa kamar koyon lissafi na quantum ko kwamfutocin hoto, kwamfutocin Neuromorphic suna ba da fa'idar dacewa ta kusa da tsarin hanyar sadarwar jijiyoyi na yanzu. Kamar yadda aka lura a cikin IEEE Transactions on Pattern Analysis and Machine Intelligence, ingancin makamashi na tsarin neuromorphic ya sa su dace musamman don aikace-aikacen AI koyaushe da na'urorin IoT. Haɗa kai tare da ƙarfin lokaci a cikin SNNs kuma yana buɗe yuwuwar sarrafa bidiyo da binciken bayanai na jeri waɗanda suka wuce dawo da hoto mai tsayi.

Ci gaban gaba zai iya bincika gine-ginen gauraye waɗanda ke haɗa ƙarfin koyon zurfi na al'ada tare da ingancin neuromorphic, kama da hanyoyin da aka tattauna a cikin Nature Machine Intelligence. Girman waɗannan tsarin zuwa manyan bayanai da ƙarin rikitattun ayyukan dawo da su ya kasance muhimmin alkibla na bincike, kamar yadda haɓaka algorithms na horo na musamman waɗanda ke daidaitawa kai tsaye don kayan aikin neuromorphic maimakon dogaro da canjin ANN-zuwa-SNN.

8. Bayanan Kafa

1. Liu, T.-Y., et al. "Neuromorphic Computing for Content-based Image Retrieval." arXiv:2008.01380 (2021)
2. Davies, M., et al. "Loihi: A Neuromorphic Manycore Processor with On-Chip Learning." IEEE Micro (2018)
3. Maass, W. "Networks of spiking neurons: The third generation of neural network models." Neural Networks (1997)
4. Roy, K., et al. "Towards Spike-based Machine Intelligence with Neuromorphic Computing." Nature (2019)
5. Xiao, H., et al. "Fashion-MNIST: A Novel Image Dataset for Benchmarking Machine Learning Algorithms." arXiv:1708.07747 (2017)
6. Merolla, P. A., et al. "A million spiking-neuron integrated circuit with a scalable communication network and interface." Science (2014)