Abstract
Standard meta-learning algorithms typically assume that training and test tasks are drawn from the same distribution. However, in real-world applications, models often encounter "meta-environment shifts" where deployment tasks differ significantly from training tasks. This paper introduces Weighted Free Energy Minimization (WFEM), a Bayesian framework for transfer meta-learning. By optimizing a weighted loss function that balances a large source-environment dataset with a limited target-environment dataset, WFEM meta-learns robust Gaussian Process (GP) priors. Our results on regression and cross-domain classification benchmarks demonstrate that WFEM effectively bridges the performance gap in out-of-distribution meta-testing scenarios.
Introduction
Meta-learning, or "learning to learn," enables models to adapt rapidly to new tasks. A critical limitation of current methods is the assumption of independent and identically distributed (i.i.d.) task environments. When a model trained on a source environment (e.g., indoor navigation) is deployed in a target environment (e.g., outdoor navigation), performance often degrades. This paper addresses the challenge of Transfer Meta-Learning: leveraging abundant source data and sparse target data to perform well on unseen tasks within the target domain.
Methodology
The proposed solution utilizes Gaussian Processes (GPs) to provide principled uncertainty quantification. The core contribution is the Weighted Free Energy Minimization (WFEM) framework, which consists of several key components:
- Weighted Meta-Loss: The objective function is defined as a convex combination of tasks from the source and target environments, governed by a weight parameter $\alpha \in [0, 1]$.
- Free Energy Objective: The framework minimizes a free energy functional that balances empirical meta-training loss with a Kullback-Leibler (KL) divergence regularization term to keep the learned hyper-posterior close to a prior.
- Meta-Learned Priors: Unlike standard GPs with fixed hyperparameters, WFEM meta-learns the mean and kernel functions of the GP prior from the available data.
- Inference Schemes: Two implementations are explored: a point-estimate Maximum A Posteriori (MAP) approach and a particle-based Stein Variational Gradient Descent (SVGD) approach, the latter of which better captures the full hyperparameter distribution.
Experimental Evaluation
WFEM-GP was evaluated on two benchmarks: a synthetic sinusoidal regression task with varying phase/offset shifts, and a cross-domain classification task transferring knowledge from mini-ImageNet (general images) to CUB (specialized bird classification).
In regression experiments, WFEM-GP significantly outperformed standard meta-learning baselines that relied solely on source data. As shown in the figure above, as the fraction of target tasks $\beta$ increases, WFEM-GP (blue line) maintains a lower Test RMSE compared to the source-only baseline (black line), effectively approaching the performance of a model trained exclusively on the target environment (green line).
For cross-domain classification, the choice of the weighting parameter $\alpha$ proved critical. Results indicate a "sweet spot" (typically $\alpha \approx 0.1$ to $0.2$) where the model achieves peak accuracy by appropriately valuing the limited target data without discarding the general knowledge from the source domain. Furthermore, the SVGD-based inference consistently outperformed the MAP estimate, highlighting the benefits of a full Bayesian treatment for hyperparameter learning.
Conclusion
WFEM provides a robust framework for transfer meta-learning by explicitly accounting for meta-environment shifts through weighted optimization. While the method requires some target-domain data during meta-training, it offers a significant performance boost over standard meta-learning. Future work may focus on automating the selection of the weighting parameter $\alpha$ and reducing the computational overhead of particle-based inference for large-scale applications.
Summary updated