Spiking neural networks (SNNs) have emerged as a promising computational model inspired by the information processing capabilities of the human brain. Unlike traditional artificial neural networks, which rely on continuous activation values, SNNs communicate through discrete time events called spikes. Such temporal nature allows SNNs to capture the dynamics of neural activity and process information in a more biologically plausible manner. The analogue realizations of spiking neural networks using complementary metal-oxide-semiconductor (CMOS) technology has gained significant attention due to the potential energy efficiency and real-time information processing built into the standard technology. Previously, a neuromorphic chip was designed and implemented in standard CMOS technology. In this Thesis, various schemes for encoding and decoding information in spikes will be explored. After examining various supervised learning algorithms tailored for SNNs, a suitable algorithm for the realized CMOS technology is proposed. Tests proved that the algorithm is capable of training SNNs but with a reduced accuracy with respect to state-of-the-art algorithms, this is traded with a simpler algorithm that can be easily implemented in an analog SNNs.
Le reti neurali spiking (SNN) sono emerse come un promettente modello computazionale ispirato alle capacità di elaborazione delle informazioni del cervello umano. A differenza delle reti neurali artificiali tradizionali, che si basano su valori di attivazione continui, le SNN comunicano attraverso eventi temporali discreti chiamati spikes. Questa natura temporale consente alle SNN di catturare le dinamiche dell'attività neurale e di elaborare le informazioni in modo biologicamente più plausibile. La realizzazione analogica di reti neurali spiking (SNN) utilizzando la tecnologia CMOS (complementary metal-oxide-semiconductor) ha guadagnato un'attenzione significativa grazie alla potenziale efficienza energetica e all'elaborazione delle informazioni in tempo reale integrate nella tecnologia standard. In un precedente lavoro di tesi, è stato progettato e implementato un chip neuromorfico in tecnologia CMOS standard. In questa tesi verranno esplorati vari schemi per la codifica e la decodifica delle informazioni negli spike. Dopo aver esaminato vari algoritmi di apprendimento supervisionato su misura per le SNN, viene proposto un algoritmo adatto alla tecnologia CMOS realizzata. I test hanno dimostrato che l'algoritmo è in grado di addestrare reti neurali di tipo spiking, ma con un'accuratezza ridotta rispetto agli algoritmi stato dell'arte; però l'algoritmo è più semplice e può essere facilmente implementato in una SNN analogica.
Supervised learning algorithms for analog spiking neural networks implemented with CMOS technology
Baschieri, Matteo
2022/2023
Abstract
Spiking neural networks (SNNs) have emerged as a promising computational model inspired by the information processing capabilities of the human brain. Unlike traditional artificial neural networks, which rely on continuous activation values, SNNs communicate through discrete time events called spikes. Such temporal nature allows SNNs to capture the dynamics of neural activity and process information in a more biologically plausible manner. The analogue realizations of spiking neural networks using complementary metal-oxide-semiconductor (CMOS) technology has gained significant attention due to the potential energy efficiency and real-time information processing built into the standard technology. Previously, a neuromorphic chip was designed and implemented in standard CMOS technology. In this Thesis, various schemes for encoding and decoding information in spikes will be explored. After examining various supervised learning algorithms tailored for SNNs, a suitable algorithm for the realized CMOS technology is proposed. Tests proved that the algorithm is capable of training SNNs but with a reduced accuracy with respect to state-of-the-art algorithms, this is traded with a simpler algorithm that can be easily implemented in an analog SNNs.File | Dimensione | Formato | |
---|---|---|---|
Supervised learning algorithms for analog spiking neural networks .pdf
accessibile in internet solo dagli utenti autorizzati
Dimensione
15.21 MB
Formato
Adobe PDF
|
15.21 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary_Supervised learning algorithms for analog spiking neural networks.pdf
accessibile in internet per tutti
Dimensione
681.24 kB
Formato
Adobe PDF
|
681.24 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/212438