Thesis title: Efficiency in Deep Learning: from Theoretical Foundations to Real-World Applications
This Thesis explores the efficiency of neural network architectures in computer vision tasks, particularly for embedded devices with limited computational resources. The study examines key components, such as self-attention, knowledge distillation, and optimizers, in order to improve learning processes in a computationally efficient manner. The Thesis addresses both perception and security tasks by focusing on practical applications such as monocular depth estimation, 3D mesh reconstruction, and deepfake detection. In more detail, this Thesis will investigate both primary areas of neural network efficiency and secondary challenges that emerge during research. Overall, it aims to provide insights into improving neural network architectures and offer guidance for advancing research in this field.