Nevronske mreže z vzvratnim razširjanjem napak v funkcijskem programskem jeziku

delo diplomskega seminarja

Laura Guzelj Blatnik (Author), Ljupčo Todorovski (Mentor), Aljaž Osojnik (Co-mentor)

Abstract

Diplomsko delo se poglobi v usmerjene nevronske mreže. Le te temeljijo na posnemanju možganskih funkcij, uporabljajo pa se za napovedovanje in klasifikacijo. Sestavljajo jih nevroni organizirani v sloje. Nevroni so med sabo povezani s sinapsami. Usmerjene nevronske mreže iz stanja nevronov v vhodnem sloju preko uteži na sinapsah in nevronih v vmesnih slojih, izračunajo napoved v izhodnem sloju. S pomočjo algoritma za vzvratno razširjanje napake in učnih primerov mrežo naučimo odzivanja na neznane situacije. Algoritem temelji na spreminjanju uteži na sinapsah, učenje pa poteka dokler ni razlika med želeno in izračunano vrednostjo dovolj majhna. Poleg nevronskih mrež se delo osredotoči tudi na funkcijsko programiranje s poudarkom na programskem jeziku OCaml. Primer nevronske mreže z vzvratnim razširjanjem napak je tudi implementiran v programskem jeziku OCaml. Delovanje mreže je prikazano na konkretnem primeru, kjer je ovrednotena napaka mreže na izbrani podatkovni množici.

Keywords

usmerjene nevronske mreže;vzvratno razširjanje napake;perceptron;umetna inteligenca;strojno učenje;funkcijski programski jezik OCaml;

Data

Language:	Slovenian
Year of publishing:	2020
Typology:	2.11 - Undergraduate Thesis
Organization:	UL FU - Faculty of Administration
Publisher:	[L. Guzelj Blatnik]
UDC:	004.8
COBISS:	58750723
Views:	1641
Downloads:	228
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Feed-forward neural networks with backpropragtion in a functional programming language
Secondary abstract:	This paper focuses on feed-forward neural networks. Mimicking brain functions, neural networks are used for prediction and classification. Neural networks are composed of neurons organized in layers and connected by synapses. From the input values through weights on the synapses and neurons in hidden layers, neural networks compute the prediction in the output layer. Utilizing the backpropagation algorithm and learning examples the network is able to learn how to respond to unknown situations. The idea behind backpropagation is to change weights on the synapses until the difference between computed and desired values is small enough. In addition, the work presents functional programming and the OCaml functional programming language. An implementation of a neural network with backpropagation in the OCaml programming language is presented at the end of the work. The work concludes with an application of the neural network to a real-world example, where the error of the network is evaluated on a specific data set.
Secondary keywords:	feed-forward neural network;backpropagation;perceptron;artificial intelligence;machine learning;functional programming language OCaml;
Type (COBISS):	Final seminar paper
Study programme:	0
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. v Ljubljani, Fak. za matematiko in fiziko, Oddelek za matematiko, Matematika - 1. stopnja
Pages:	38 str.
ID:	11907164