Παρακαλώ χρησιμοποιήστε αυτό το αναγνωριστικό για να παραπέμψετε ή να δημιουργήσετε σύνδεσμο προς αυτό το τεκμήριο:
|Τίτλος:||Functional assessment of retinal models using machine learning techniques / Λειτουργική αξιολόγηση μοντέλων αμφιβληστροειδούς με χρήση τεχνικών μηχανικής μάθησης|
|Λέξεις κλειδιά:||retinal models|
|Περίληψη:||This diploma thesis proposes the functional assessment of retinal models, as an alternative to the currently common practice of comparing the similarity of model-generated and ground truth neural responses, as recorded by retinal implants. Functional assessment describes the concept of evaluating the performance of retinal models on image understanding tasks. In this work, we developed a pipeline for functional assessment using machine learning techniques, where we feed retinal models with images and we receive the neural responses of the model, with which we train classifiers on object and digit recognition tasks. In particular, we used CIFAR10, Fashion MNIST and Imagenette datasets for object recognition tasks and MNIST dataset for digit recognition tasks. We also trained common classifiers, such as Multilayer Perceptron (MLP), Support Vector Machine (SVM) and Random Forest. We examined a number of relevant questions: which classifiers perform better in such type of tasks, how the performance of a model differs depending on the task and how the number of available neurons affects the performance of the model. Given also that the Convolutional Neural Network (CNN) retinal model provides a limited number of available neurons (60 in our case), we investigated ways to artificially increase this number. For that, we split the images into different parts, we fed separately each part to the retinal model and we investigated how we can optimally combine the neural responses produced by the model for each part, in order to achieve high performance in image recognition tasks. We found that Random Forest classifier achieves the highest performance and that models perform better with datasets such as MNIST and Fashion MNIST, where we achieved up to 90% accuracy. In addition, simulations indicated that using more neurons improves the overall performance of the model, thus highlighting the need to design retinal prostheses with a larger number of available neurons. Finally, we applied functional assessment in order to compare the performance of two different retinal models. The results showed that there is an agreement between standard and functional assessment, as the two models had similar performance in both techniques. Thus, we conclude that functional assessment produces reliable results and it can be used as an alternative to the standard evaluation technique.|
|Εμφανίζεται στις συλλογές:||Διπλωματικές Εργασίες - Theses|
Αρχεία σε αυτό το τεκμήριο:
|Thesis-ECE_NTUA-Nikolas_Papadopoulos-EN_GR.pdf||2.51 MB||Adobe PDF||Εμφάνιση/Άνοιγμα|
Όλα τα τεκμήρια του δικτυακού τόπου προστατεύονται από πνευματικά δικαιώματα.