Date of Publication

2024

Document Type

Dissertation/Thesis

Degree Name

Master of Science in Computer Science

College

College of Computer Studies

Department/Unit

Software Technology

Thesis Advisor

Joel P. Ilao

Macario O. Cordel II

Defense Panel Chair

Ann Franchesca B. Laguna

Defense Panel Member

Judith J. Azcarraga

Abstract (English)

Object detectors are used in a wide range of different applications especially in critical situations such as autonomous driving. These models are trained in a closed environment and a balanced dataset and when they are deployed in the real-world they will encounter unknown classes. However, when traditionally updating the model with new classes it results in catastrophic forgetting the previous classes. Strategies such as replay and knowledge distillation alleviate this issue and use the model’s accuracy to manage its training. Instead, in our proposed method we guide the model’s training through a post-hoc interpretability module. We use the network dissection to visualize the knowledge of the backbone of the object detector when there is a significant change in model performance. Furthermore, when the change of knowledge visualization is substantial, we feed the current model’s training with the previous classes. Our results show that our method a low performance compared to the replay strategy when training on a small number of classes. However, when the model has been trained on a large number of previous classes our method will guide the model’s training better by determining specific previous classes to feed in the current model training.

Abstract Format

html

Language

English

Upload Full Text

wf_yes

Embargo Period

4-23-2024

Download

Share

COinS