Publicação
Intelligent Model Adaptation: A Study of the Application of Deep Neural Networks Repair Techniques in Transfer Learning
| Resumo: | Deep Neural Networks (DNNs) are widely used across various applications but often exhibit unexpected behaviours. Current methods, including systematic retraining, have limitations, such as resource intensity and a lack of control over localized changes. Transfer Learning (TL) offers improvements but introduces its own challenges, including domain adaptation and bias correction. Search-based repair techniques are presented as a potential solution to mitigate these issues. This work explores the integration of search-based repair techniques into TL for DNNs, a novel approach with untapped potential. The main objective is to enhance TL efficiency and accuracy. To achieve this, a new framework that combines the search-based repair technique Arachne with TL, named IMA (Intelligent Model Adaptation), is developed and benchmarked against traditional TL methods using the Cifar-10 dataset. Overall, the preliminary results reveal that integrating DNN repair techniques has the potential to achieve comparable or enhanced TL performance, whilst highlighting the need for further experimentation and framework optimization. |
|---|---|
| Autores principais: | Velho, Mafalda Salvado Gouveia De Sá |
| Assunto: | Deep learning Transfer learning DNN repair Search-based repair techniques Arachne Aprendizagem profunda Aprendizagem por transferência Reparação de redes neuronais profundas Ténicas de reparação baseadas em pesquisa Arachne SDG 9 - Industry, innovation and infrastructure |
| Ano: | 2024 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Deep Neural Networks (DNNs) are widely used across various applications but often exhibit unexpected behaviours. Current methods, including systematic retraining, have limitations, such as resource intensity and a lack of control over localized changes. Transfer Learning (TL) offers improvements but introduces its own challenges, including domain adaptation and bias correction. Search-based repair techniques are presented as a potential solution to mitigate these issues. This work explores the integration of search-based repair techniques into TL for DNNs, a novel approach with untapped potential. The main objective is to enhance TL efficiency and accuracy. To achieve this, a new framework that combines the search-based repair technique Arachne with TL, named IMA (Intelligent Model Adaptation), is developed and benchmarked against traditional TL methods using the Cifar-10 dataset. Overall, the preliminary results reveal that integrating DNN repair techniques has the potential to achieve comparable or enhanced TL performance, whilst highlighting the need for further experimentation and framework optimization. |
|---|