Optimierung der Konvertierung kontinuierlich aktivierter Netzwerke zu gepulsten neuronalen Netzwerken
Year: 2023
Type: dissertation
Abstract: We present optimization techniques and novel approaches for the conversion of continuous-valued networks to spiking neural networks (SNNs). This enables faster inference and improved hyperparameter approximation in rate-coded networks providing a basis for very deep SNNs and for spiking transformer networks. Furthermore, it allows for sparse, low-energy computation with temporal-coded networks and previously infeasible time-series processing through population coding.
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7032983574