Curiosity-driven AI for Science: Automated Discovery of Self-Organized Structures. (IA Curieuse au service de la Science: Découverte Automatisée de Structures Auto-Organisées)

Mayalen Etcheverry

Abstract: Complex systems are very hard to predict and control due to their chaotic dynamics and open-ended outcomes. However, understanding and harnessing the underlying mechanisms of these systems hold great promise for revolutionizing many areas of science. While considerable progress has been made in manipulating and measuring system activity down to the lowest level, there remains a fundamental gap between our knowledge at the micro-level and our ability to control resulting properties on a global scale. Modern machine learning tools offer promising avenues for assisting scientists in navigating the vast space of possible outcomes, especially when aiming for novel or challenging morphological or functional objectives. Nevertheless, current methods tend to constrain and bias the range of events that AI can measure and attempt to influence. This thesis aims to transpose and advance recent computational models of intrinsically motivated learning and exploration with the goal of designing more open-ended forms of AI discovery assistants for assisting scientists in mapping the outcome space of self-organizing systems. To that end, several key ingredients are introduced to efficiently shape the discovery process. These include the use of unsupervised learning for representations, meta-diversity search, curriculum learning, and external human guidance, whether environment-based or preference-based. We discuss how these components, when implemented in practice, can help address challenging problems in science. These challenges encompass the search for interesting patterns in continuous models of cellular automata, the investigation of the origins of sensorimotor agency, the exploration of gene regulatory networks behavioral capabilities, and the design of innovative forms of cellular collectives for applications in AI and biology.