delo diplomskega seminarja
Tim Dolenc (Author), Ljupčo Todorovski (Mentor)

Abstract

Diplomska naloga raziskuje usmerjene nevronske mreže in rekurenčne nevronske mreže (RNN). Predstavljene so osnovne značilnosti, algoritmi za učenje obeh vrst mrež in dodatni način predstavitve gradienta funkcije izgube pri RNN. Delo vsebuje tudi rezultate lastne implementacije usmerjene nevronske mreže in aplikacijo RNN na problem napovedovanja porabe elektrike. Rezultati kažejo, da so RNN primerne za kratkoročno napovedovanje časovnih vrst, pri daljših zaporedjih pa se soočajo z izzivi kot sta eksplodirajoči in izginjajoči gradient.

Keywords

nevronske mreže;usmerjene nevronske mreže;rekurenčne nevronske mreže;napovedovanje časovnih vrst;gradient funkcije izgube;verižno pravilo;vzvratno razširjanje napake;

Data

Language: Slovenian
Year of publishing:
Typology: 2.11 - Undergraduate Thesis
Organization: UL FMF - Faculty of Mathematics and Physics
Publisher: [T. Dolenc]
UDC: 004.8
COBISS: 206082819 Link will open in a new window
Views: 34
Downloads: 11
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Reccurent neural networks and their applications
Secondary abstract: The thesis explores feedforward neural networks and recurrent neural networks (RNNs). It presents the fundamental characteristics and learning algorithms for both types of networks and introduces an additional method for representing the loss function gradient in RNNs. The work also includes a custom implementation of a feedforward neural network and the application of RNNs to the problem of forecasting electricity consumption. The results indicate that RNNs are suitable for short-term time series forecasting, although they face challenges such as exploding and vanishing gradients.
Secondary keywords: neural networks;feedforward neural networks;recurrent neural networks;time series forecasting;loss function gradient;chain rule;backpropagation;
Type (COBISS): Final seminar paper
Study programme: 0
Thesis comment: Univ. v Ljubljani, Fak. za matematiko in fiziko, Oddelek za matematiko, Finančna matematika - 1. stopnja
Pages: 27 str.
ID: 24870429