|
|
Design of Accuracy Predictors for Convolutional Neural Networks
Šmída, Šimon ; Mrázek, Vojtěch (oponent) ; Sekanina, Lukáš (vedoucí práce)
The aim of this thesis is to present a method of constructing accuracy predictors for convolutional neural networks (CNNs) by leveraging databases of trained CNNs (NAS-Bench-101) and employing machine learning (ML) techniques as performance estimation strategies. The study begins with a description of various ML methods used in building CNN accuracy predictors, followed by an in-depth examination of CNNs and databases of pre-trained CNNs. The proposed method involves selecting a suitable task for the CNNs (image classification), assembling a dataset, defining relevant features for the predictor input, and choosing five ML methods for training the predictors. Using existing libraries, the accuracy predictors are implemented, trained, and experimentally validated to assess their functionality and performance. The results are thoroughly evaluated, providing insights into the effectiveness of the proposed method and the potential for further refinement in the field of CNN accuracy prediction.
|
|
|
Optimizing neural network architecture for EEG processing using evolutionary algorithms
Pijáčková, Kristýna ; Maršálek, Roman (oponent) ; Götthans, Tomáš (vedoucí práce)
This thesis deals with an optimization of neural network hyperparameters for EEG signal processing using evolutionary algorithms. The incorporation of evolutionary optimization can reduce reliance on human intuition and empirical knowledge when designing neural network and can thus make the process design more effective. In this work, a genetic algorithm was proposed that is suitable for hyperparameters optimization as well as neural architecture search. These methods were compared to a benchmark model designed by an engineer with expertise in iEEG processing. Data used in this work are classified into four categories and come from St. Anne's University Hospital (SAUH) and Mayo Clinic (MAYO) and were recorded on drug-resistant epileptic patients undergoing pre-surgical examination. The results of the neural architecture search method were comparable with the benchmark model. The hyperparameter optimization improved the F1 score over the original, empirically designed, model from 0.9076 to 0.9673 for the SAUH data and 0.9222 to 0.9400 for the Mayo Clinic data. The increased scores were mainly due to the increased accuracy of the classification of pathological events and noise, which may have further positive implications in applications of this model in seizure and noise detectors.
|
host ::
RSS.