David Rohrschneider

• AI research and applications have experienced a paradigm shift with the emergence of foundation models. Although these models deliver state-of-the-art results across a broad spectrum of tasks, their exponential increase in size necessitates methods for parameter­efficient training. To tackle this, the integration of adapter modules represents a promising yet compact strategy, offering remarkable performance by incorporating a minimal num­ber of parameters for each specific task. This research proposes the "repetitive adapter module", designed to add a layer of linear scalability to the traditional bottleneck ar­chitecture. By integrating these modules into the foundational models SEEM and Mask DINO and applying them to three downstream tasks, the approach is demonstrated nearly reaching the effectiveness of traditional fine-tuning methods while requiring significantly fewer trainable parameters. Furthermore, the thesis highlights the applicability of repet­itive adapters within the modular meta-architecture underlying SEEM and Mask DINO, proving their effectiveness disregarding model size and multi-modality. This investiga­tion also explores the limitations of adapter fine-tuning, laying the groundwork for future research in this domain.