Maria Matveev

Brief description of my doctoral project

In my research, I investigate why artificial neural networks tend to learn general patterns instead of simply memorizing data, despite their capacity to do so. I focus on the “implicit bias” of training algorithms and study how the training process itself leads to generalization. Methodologically, I combine theoretical analyses of simplified models, enabling precise mathematical insights, with large-scale numerical experiments on realistic models, including LLMs. A deeper understanding of these internal mechanisms aims to improve the reliability and trustworthiness of AI systems and to inform the design of more robust models and learning methods.