Deep learning algorithms with emphasis on deep convolutional neural networks for optimization of bronchodilator inhaler’s use

Nowadays, artificial intelligence and machine learning algorithms have seen an enormous growth of interest with several researchers dedicating their lives into seeking new algorithms or improving those already developed. Studies have shown that they can be applied in almost every system that is r...

Πλήρης περιγραφή

Λεπτομέρειες βιβλιογραφικής εγγραφής
Κύριος συγγραφέας: Νταλιάνης, Ευάγγελος
Άλλοι συγγραφείς: Μπίρμπας, Μιχαήλ
Μορφή: Thesis
Γλώσσα:English
Έκδοση: 2020
Θέματα:
Διαθέσιμο Online:http://hdl.handle.net/10889/12980
Περιγραφή
Περίληψη:Nowadays, artificial intelligence and machine learning algorithms have seen an enormous growth of interest with several researchers dedicating their lives into seeking new algorithms or improving those already developed. Studies have shown that they can be applied in almost every system that is required to make a decision. Machine and deep learning algorithms are widely used in image or sound classification tasks, in biomedical applications and language processing. In this thesis we study the performance of Convolutional Neural Network in an audio classification task. In particular, our classifier is fed with real time data which are produced by recording the use of an inhaler device and it attempts to categorize them into four classes, which namely are inhalation, exhalation, drug and noise and other environmental sounds. Furthermore, we study the potential of pruning the weights and how zeroing out weights with small magnitude affects the overall performance of the classifier. So, firstly, we describe the most common respiratory diseases and why it is essential to monitor the medication adherence. Secondly, we introduce to the reader the fundamentals of machine learning as well as the most important characteristics of algorithms used for the problem of monitoring the medication adherence in respiratory diseases through audio classification. Finally, we propose different approaches for compressing our model and we present several metrics in order to evaluate the performance of our classifier both before and after pruning procedure is employed.