TLXML: Task-Level Explanation of Meta-Learning via Influence Functions

TL;DR

TLXML tackles explainability in meta-learning by quantifying how prior training tasks shape adaptation and inference through task-level influence functions. It adapts influence-function theory to a bi-level meta-learning setting and employs a Gauss-Newton Hessian approximation to dramatically reduce computational cost, while handling flat directions with a Hessian pseudo-inverse. Empirical results show TLXML can distinguish between training vs. test tasks and between task distributions (e.g., clean versus noisy data) for MAML and Prototypical Networks, providing a concise similarity metric between training and test tasks. This work advances safe, interpretable meta-learning with scalable explanations and sets the stage for extensions to regression, reinforcement learning, and broader XAI applications.

Abstract

The scheme of adaptation via meta-learning is seen as an ingredient for solving the problem of data shortage or distribution shift in real-world applications, but it also brings the new risk of inappropriate updates of the model in the user environment, which increases the demand for explainability. Among the various types of XAI methods, establishing a method of explanation based on past experience in meta-learning requires special consideration due to its bi-level structure of training, which has been left unexplored. In this work, we propose influence functions for explaining meta-learning that measure the sensitivities of training tasks to adaptation and inference. We also argue that the approximation of the Hessian using the Gauss-Newton matrix resolves computational barriers peculiar to meta-learning. We demonstrate the adequacy of the method through experiments on task distinction and task distribution distinction using image classification tasks with MAML and Prototypical Network.

Figures (9)

• Figure 1: Without TLXML, it is common to explain the model's behavior via local explanations.
• Figure 2: TLXML calculates the influence of each previously learned task on the model's behavior in a given new task, resulting in a more effective user explanation.
• Figure 4: Diagram of the projected influence function, which measures the influence of a training task on the meta-parameters with the Hessian flat directions projected out.
• Figure 5: Test with training tasks. As shown in (a), the task most similar to the test task is successfully separated from the others by using TLXML.
• Figure 6: Test with degraded training tasks. The parameters $\alpha$ and ratio specify the darkness of images, and the proportion of the dark images, respectively. The red and blue lines represent ranks and scores, respectively. Both examples were performed with the Hessian pruned to retain only the 1193 most significant eigenvalues.