Jahr | 2015 |
Autor(en) | Roman Martel |
Titel | Generative Properties of LIF-based Boltzmann Machines |
KIP-Nummer | HD-KIP 15-86 |
KIP-Gruppe(n) | F9 |
Dokumentart | Masterarbeit |
Abstract (de) | Boltzmann-Maschinen sind künstliche neuronale Netze, die in der Lage sind generative Modelle realistischer Daten zu lernen. Variationen von Boltzmann-Maschinen werden derzeit erfolgreich in vielen Problemstellungen des maschinellen Lernens angewendet. |
Abstract (en) | Boltzmann machines are artificial neural networks which are able to learn generative models of real world data. Variations of Boltzmann machines are currently successfully applied to many machine learning tasks. The generation of samples (sampling) from a learned model in Boltzmann machines is computationally expensive but, in principle, massively parallel, which makes it appealing for an implementation on neuromorphic computing platforms. This has motivated the development of Boltzmann machines based on the leaky integrate-and-fire (LIF) neuron model, which is often implemented on neuromorphic devices. This thesis investigates the sampling properties of LIF-based Boltzmann machines compared to the one of classical Boltzmann machines. We observe differences in sampling, especially with regard to the ability to mix between several modes of the underlying model. Depending on the synapse model, connecting the neurons, an improvement or a decline of the mixing ability compared to classical sampling is found. Moreover, we present a mechanism based on short-term synaptic plasticity to improve the mixing ability for LIF-based Boltzmann machines. Using this mechanism, a better mixing ability compared to classical sampling for both applied synapse models is achieved. |
bibtex | @mastersthesis{martel2015masterthesis, author = {Roman Martel}, title = {Generative Properties of LIF-based Boltzmann Machines}, school = {Ruprecht-Karls-Universität Heidelberg}, year = {2015}, type = {Master thesis}, note = {HD-KIP 15-86} } |
Datei |