Efficiënte methoden voor afstemming en generatie in visie en taal

In the modern era of deep learning with neural networks, the overarching goal is to develop models that can perceive the physical world through vision and communicate effectively through language, much like humans. Achieving this objective typically involves pre-training models with a large number of parameters on extensive datasets, enabling precise alignment and accurate generation when adapting them to downstream tasks. However, the considerable resource demands necessitate the development of efficient methods for adaptation. We focus on efficiently reusing pre-trained models while preserving their inherent knowledge to the greatest extent possible. This dissertation focuses on four key efficiency challenges: parameter efficiency, data efficiency, training efficiency, and inference efficiency. Addressing these challenges, we primarily concentrate on alignment and generation problems in Vision-Language (VL) tasks, as vision and language permeate our daily lives. We also extend our work to generative models, which are naturally applicable to vision and language.