[Paper Review] Recasting Gradient-Based Meta-Learning as Hierarchical Bayes

Erin Grant, Chelsea Finn+3 more|arXiv (Cornell University)|Jan 26, 2018

Gaussian Processes and Bayesian Inference36 references167 citations

TL;DR

The paper shows that gradient-based meta-learning (MAML) can be interpreted as Bayesian hierarchical inference, linking MAML to empirical Bayes, and proposes Bayesian-inspired improvements to enhance performance and scalability.

ABSTRACT

Meta-learning allows an intelligent agent to leverage prior learning episodes as a basis for quickly improving performance on a novel task. Bayesian hierarchical modeling provides a theoretical framework for formalizing meta-learning as inference for a set of parameters that are shared across tasks. Here, we reformulate the model-agnostic meta-learning algorithm (MAML) of Finn et al. (2017) as a method for probabilistic inference in a hierarchical Bayesian model. In contrast to prior methods for meta-learning via hierarchical Bayes, MAML is naturally applicable to complex function approximators through its use of a scalable gradient descent procedure for posterior inference. Furthermore, the identification of MAML as hierarchical Bayes provides a way to understand the algorithm's operation as a meta-learning procedure, as well as an opportunity to make use of computational strategies for efficient inference. We use this opportunity to propose an improvement to the MAML algorithm that makes use of techniques from approximate inference and curvature estimation.

Motivation & Objective

• Motivate fast adaptation across related tasks using prior learning episodes.
• Formalize meta-learning as inference in a hierarchical Bayesian model.
• Connect MAML to empirical Bayes and Bayesian posterior estimation to provide theoretical grounding.
• Propose enhancements to MAML using approximate inference and curvature-aware updates.

Proposed method

• Reformulate MAML as inference in a hierarchical Bayesian model with a shared meta-parameter theta and task-specific phi_j.
• Show that MAML’s inner-loop gradient steps correspond to a prior over phi_j conditioned on theta.
• Derive the equivalence between MAML objective and marginal likelihood maximization with respect to theta using a point estimate for phi_j.
• Propose improvements via Laplace approximation and curvature-based preconditioning to better approximate the marginal likelihood.
• Discuss the role of early stopping as inducing a prior over task-specific parameters.
• Provide algorithmic variants (e.g., MAML-HB) that integrate Bayesian ideas into the meta-learning loop.

Experimental results

Research questions

• RQ1 Can gradient-based meta-learning be interpreted as Bayesian hierarchical inference?
• RQ2 How does viewing MAML through a Bayesian lens illuminate the algorithm’s successes and limitations?
• RQ3 What Bayesian-inspired modifications can improve MAML’s performance and stability?

Key findings

• MAML can be understood as empirical Bayes performing marginal likelihood optimization over a shared meta-parameter.
• The inner-loop updates act as a prior over task-specific parameters, linking task solutions through theta.
• Early stopping and truncated gradient updates correspond to implicit priors in a hierarchical Bayesian formulation.
• Laplace’s method and curvature-based updates offer principled approximations to incorporate uncertainty and improve inference.
• Bayesian reinterpretations motivate algorithmic improvements applicable to complex function approximators beyond simple linear models.