Online Hyperparameter Meta-Learning with Hypergradient Distillation
This work addresses scalability and online optimization challenges in hyperparameter tuning for meta-learning, offering a practical solution for researchers and practitioners in machine learning.
The paper tackles the limitations of existing gradient-based hyperparameter optimization methods in meta-learning, such as scalability and online optimization issues, by proposing a novel method that approximates second-order terms with knowledge distillation, achieving effective results on two meta-learning methods and three benchmark datasets.
Many gradient-based meta-learning methods assume a set of parameters that do not participate in inner-optimization, which can be considered as hyperparameters. Although such hyperparameters can be optimized using the existing gradient-based hyperparameter optimization (HO) methods, they suffer from the following issues. Unrolled differentiation methods do not scale well to high-dimensional hyperparameters or horizon length, Implicit Function Theorem (IFT) based methods are restrictive for online optimization, and short horizon approximations suffer from short horizon bias. In this work, we propose a novel HO method that can overcome these limitations, by approximating the second-order term with knowledge distillation. Specifically, we parameterize a single Jacobian-vector product (JVP) for each HO step and minimize the distance from the true second-order term. Our method allows online optimization and also is scalable to the hyperparameter dimension and the horizon length. We demonstrate the effectiveness of our method on two different meta-learning methods and three benchmark datasets.