UC Irvine

Autonomous vehicles rely on machine learning models that must make accurate predictions under strict latency, memory, energy, and reliability constraints. However, modern deep learning systems are often designed as fixed, dense inference pipelines, even though real driving conditions vary across compute availability, scene difficulty, sensor structure, and modality access. This dissertation presents adaptive sparsity as a framework for making autonomous vehicle inference more efficient and responsive to these changing conditions. This dissertation explores three complementary forms of sparsity. First, parameter sparsity enables dynamic adjustment of model capacity in response to runtime system constraints. Second, activation sparsity exploits input-dependent computation to reduce unnecessary work during inference. Third, input sparsity uses rich training-time modalities to improve deployment-time models that operate with fewer or restricted inputs. Together, these methods show that sparsity can serve not only as a compression tool, but also as a mechanism for adaptive, efficient, and reliable inference in autonomous vehicle systems.