Guiding AI Attention for Driving and Creative Generation

Diego Porres Bustamante successfully defended his dissertation on Computer Science on December 5, 2024, and he is now Doctor of Philosophy by the Universitat Autònoma de Barcelona.

What is the thesis about?

Artificial Intelligence (AI) has accelerated the advancement of numerous fields, particularly where data and computational resources are readily accessible. This thesis explores two of these domains: autonomous driving and visual arts, focusing on enhancing efficiency, interpretability, and creative potential through innovative applications of attention mechanisms and visual perception.

In the realm of autonomous driving, we address the challenge of training models in a more data-efficient manner while avoiding additional computational overhead during deployment. We introduce a novel Attention Loss that optimizes attention weights in end-to-end models, significantly improving sample efficiency and interpretability. This approach eliminates the need for dedicated masking networks, reducing computational requirements while maintaining performance. Furthermore, we investigate the stability of driving quality in end-to-end models, proposing a new loss function that separates attention into dynamic objects and traffic rules, enhancing the transparency of decision-making processes and paving the way for integrating human gaze data.

Concurrently, we tackle the challenges posed by generative AI in visual arts. Despite the immense data and computational requirements for training generative models to create unseen images, we observe a concerning lack of fresh data, potentially leading to catastrophic implications. Our work focuses on exploiting existing models more effectively by developing new interaction methods and repurposing model components for tasks beyond their original training. We introduce novel techniques to enhance the creative capabilities of generative models, propose methods for evaluating their impact on the artistic community, and explore innovative interfaces for human-AI collaboration in art creation.

Through extensive experiments and critical analysis, we demonstrate that our approaches advance the state-of-the-art in their respective fields and reveal unexpected synergies between autonomous systems and creative AI applications. This thesis contributes to developing more interpretable and efficient autonomous driving systems, while also pushing the boundaries of AI-assisted art creation in ways that respect and enhance human creativity.

Keynotes

Autonomous driving, computer vision, machine learning, deep learning, attention mechanisms, generative AI, visual arts, human-AI collaboration