|
|
Evolutionary Design of Neural Networks
Kastner, Jan ; Hurta, Martin (referee) ; Sekanina, Lukáš (advisor)
The thesis deals with the implementation of a problem-solving method for the automated design of convolutional neural networks (CNN) architectures. The optimization of two fundamental and often conflicting characteristics, the number of parameters and the quality of CNN classification, is performed using a multi-criteria optimization genetic algorithm (NSGA-II). To encode this problem, the Cartesian genetic programming (CGP) technique is used, which enables the wide range of CNN architectures to be represented, and at the same time, the searched area can be appropriately limited by parameterization. Experiments were performed on the MNIST dataset to understand the effect of population size on the quality of the resulting solution. It is also evident from the results of the experiments that the quality of the architectures found can compete with already established models. This is therefore an alternative approach that does not require human intervention compared to manual design.
|
|
| |
|
|
Evolutionary Design of Non-Linear Functions for Convolutional Neural Networks
Hladiš, Martin ; Mrázek, Vojtěch (referee) ; Sekanina, Lukáš (advisor)
The aim of this thesis is to design and implement a program for automated design of nonlinear activation functions for convolutional neural networks (CNN) using evolutionary algorithms. The use of automated design provides an independent view to systematically explore a wide range of activation functions and identify the best ones. The method for automatic design chosen in this thesis is a form of evolutionary algorithms referred to as Cartesian genetic programming, which uses a graph representation to encode the solution. This technique allows for the definition of a set of mathematical primitives that define the search space, and thus simply parameterize the design. The implemented approach has been tested on several different architectures and datasets (LeNet-5 \& MNIST, ResNet-10 \& FashionMNIST, WRN-40-4 \& CIFAR-10). Experiments have shown that the approach can find activation functions that statistically improve the accuracy of the architecture over the commonly used ReLU function.
|
|
|
Automated compression of neural network weights
Lorinc, Marián ; Sekanina, Lukáš (referee) ; Mrázek, Vojtěch (advisor)
Konvolučné neurónové siete (CNN) od svojho vynájdenia zrevolucionizovali spôsob, akým sa realizujú úlohy z odvetvia počítačového videnia. Vynález CNN viedol k zníženiu pamäťovej náročnosti, keďže váhy boli nahradené konvolučnými filtrami obsahujúcimi menej trénovateľných váh. Avšak, toto zníženie bolo dosiahnuté na úkor zvýšenia požiadaviek na výpočtový výkon, ktorý je naviazaný na výpočet konvolúcie. Táto práca skúma hypotézu, či je možné sa vyhnúť načítavaniu váh a miesto toho ich vypočítať, čím sa ušetrí energia. Na otestovanie tejto hypotézy bol vyvinutý nový algoritmus kompresie váh využívajúci Kartézske genetické programovanie. Tento algoritmus hľadá najoptimálnejšiu funkciu kompresie váh s cieľom zvýšiť energetickú účinnosť. Experimenty vykonané na architektúrach LeNet-5 a MobileNetV2 ukázali, že algoritmus dokáže efektívne znížiť spotrebu energie pri zachovaní vysokej presnosti modelu. Výsledky ukázali, že určité vrstvy je možné doplniť vypočítanými váhami, čo potvrdzuje potenciál pre energeticky efektívne neurónové siete.
|
|
|
Automated Representation Learning for Cartesian Genetic Programming Using Neural Networks
Koči, Martin ; Mrázek, Vojtěch (referee) ; Sekanina, Lukáš (advisor)
This master's thesis addresses the integration of neural networks and Cartesian Genetic Programming (CGP). It explores the use of neural networks for automated representation creation for CGP and their application to improve the evolutionary process in CGP. The study covers basic concepts of machine learning, including various types of learning and neural network models. It also touches on evolutionary algorithms with an emphasis on their basic principles, general algorithms, and types of representations. This work also includes principles of representation learning and two fundamental architectures for their creation. It describes the subsequent use of representation learning in genetic programming. The solution design includes data acquisition and preprocessing, representation creation processes, and the utilization of the resulting representations. The thesis also implements two new approaches for creating representations for Cartesian genetic programs. It further explores their use in two new mutation operators, where one is based on direct modification of the vector representation and the other on the selection of genes for mutation based on their similarity. The last of the explored areas is predicting the suitability of candidate solutions using newly emerged representations.
|
|
|
Srovnání metod operátoru křížení v kartézském genetickém programování
Fedorová, Ela ; Mrázek, Vojtěch (referee) ; Hurta, Martin (advisor)
This thesis deals with the crossover operator in Cartesian genetic programming. The aim of the thesis is to investigate and compare the use of different crossover methods. The thesis focuses on analyzing the behavior and effectiveness of selected methods on symbolic regression problems. My own implementation of three crossover methods, namely the subgraph method, block method and discrete recombination, was used. For these methods, experiments were performed to monitor the number of generations to find a solution and the quality of the solution found. The results indicate the potential of all three investigated methods to improve CGP performance and point to the subgraph method as the most useful in the given experimental setting.
|
|
|
Využití operátoru křížení v kartézském genetickém programování
Bromnik, Petr ; Sekanina, Lukáš (referee) ; Hurta, Martin (advisor)
The aim of this paper is to propose and implement two new crossover methods in Cartesian Genetic Programming (CGP) and compare them with existing approaches. CGP is a type of evolutionary algorithm that uses acyclic graphs to represent executable programs. Most CGP applications use the mutation operator only, but the effort to find a suitable crossover operator is still ongoing. In this paper, the two newly proposed crossover methods are compared on five symbolic regression problems against the standard 1 + lambda procedure based purely on mutation. Experimental results show that these methods find solutions in a similar number of fitness evaluations as 1 + lambda and, in two cases, even significantly earlier.
|
|
|
Evolutionary Design of Local Image FIlters
Gall, Samuel ; Hurta, Martin (referee) ; Sekanina, Lukáš (advisor)
This thesis focuses on the research and implementation of evolutionary design of local image filters. The aim is to create a tool capable of automatically designing suitable image operators for image filtering, specifically noise removal, using an evolutionary design method called Cartesian Genetic Programming (CGP). The tool was used for various experiments with different settings of CGP parameters such as grid size, population size, and mutation parameter. The created filters were compared with conventional noise removal filters. Evolved filters were tested on a set of test images, where their behavior was comparable to that of a median filter. Unlike the median filter, evolved filters were able to preserve more image quality.
|
|
|
Genetic Programming with Memory for Symbolic Regression
Jůza, Tadeáš ; Bidlo, Michal (referee) ; Sekanina, Lukáš (advisor)
The purpose of this thesis is to evaluate the possibility of extending genetic programming with memory for solving symbolic regression problems. Furthermore, a set of problems for testing the quality of such solutions is developed. The thesis proposes a practical application of such an extension to reduce the energy consumption of loading weights of convolutional neural networks. Instead of retrieving all the weights of the network from external memory, only a small percentage of the weights is retrieved and the remaining ones are generated using the evolved expression. This method was primarily evaluated on reducing the set of weights of convolutional layers of a small convolutional neural network classifying the MNIST dataset. Furthermore, the possibility of generating weights was also tested on other convolutional neural networks solving more complex classification problems. The proposed method has delivered interesting tradeoffs between the classification accuracy and weight memory size.
|
|
|
Symbolic Regression and Coevolution
Drahošová, Michaela ; Žaloudek, Luděk (referee) ; Sekanina, Lukáš (advisor)
Symbolic regression is the problem of identifying the mathematic description of a hidden system from experimental data. Symbolic regression is closely related to general machine learning. This work deals with symbolic regression and its solution based on the principle of genetic programming and coevolution. Genetic programming is the evolution based machine learning method, which automaticaly generates whole programs in the given programming language. Coevolution of fitness predictors is the optimalization method of the fitness modelling that reduces the fitness evaluation cost and frequency, while maintainig evolutionary progress. This work deals with concept and implementation of the solution of symbolic regression using coevolution of fitness predictors, and its comparison to a solution without coevolution. Experiments were performed using cartesian genetic programming.
|
guest ::
