UC Davis

Modern machine learning has advanced rapidly, but much of this progress depends on large amounts of labeled data, computation, memory, storage, and communication. These requirements make state-of-the-art systems expensive to reproduce, adapt, and deploy, limiting participation to organizations with substantial resources. Even when code and model checkpoints are publicly available, the cost of working with them can remain out of reach for many academic laboratories and smaller research groups. This dissertation studies how to reduce these barriers and make advanced machine learning more accessible.The central idea of this dissertation is that efficiency should be considered throughout the full lifecycle of a model. Learning should depend less on expensive supervision. Models and learned representations should require fewer parameters and less storage. Systems should be easier to communicate and deploy on devices with limited memory or bandwidth. Task-specific capabilities should also be reusable and easier to combine, rather than requiring repeated training of large models from scratch.To support these goals, this dissertation develops methods across visual representation learning, compact neural network parameterization, ecient 3D scene representation, and parameter-ecient model adaptation. Although these problems arise in dierent settings, they are connected by the same objective: reducing resource demands without giving up the performance and flexibility that make modern machine learning useful. The results show that eciency can often be improved by changing how models learn, how they represent information, and how they share structure across tasks.The broader goal of this work is to help move machine learning toward systems that are not only capable, but also aordable, reusable, and practical under limited resources. By lowering the cost of training, storing, deploying, and combining models, this dissertation aims to support a wider and more inclusive research community.