A Sample Complexity Separation between Non-Convex and Convex Meta-Learning

Authors: Nikunj Saunshi, Yi Zhang, Mikhail Khodak, Sanjeev Arora

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical This work shows that convex-case analysis might be insufficient to understand the success of meta-learning, and that even for non-convex models it is important to look inside the optimization black-box, specifically at properties of the optimization trajectory. We construct a simple meta-learning instance that captures the problem of one-dimensional subspace learning. For the convex formulation of linear regression on this instance, we show that the new task sample complexity of any initialization-based meta-learning algorithm is (d), where d is the input dimension. In contrast, for the non-convex formulation of a two layer linear network on the same instance, we show that both Reptile and multi-task representation learning can have new task sample complexity of O(1), demonstrating a separation from convex meta-learning. Crucially, analyses of the training dynamics of these methods reveal that they can meta-learn the correct subspace onto which the data should be projected.
Researcher Affiliation Academia 1Princeton University, Princeton, New Jersey, USA 2Carnegie Mellon University, Pittsburgh, Pennsylvania, USA 3Institute for Advanced Study, Princeton, New Jersey, USA.
Pseudocode No The paper describes algorithms in prose (e.g., "Reptile: Starting from (A0, w0), the initialization maintained by the algorithm is sequentially updated...") but does not include formal pseudocode blocks or labeled algorithm figures.
Open Source Code No The paper does not contain any statement about releasing source code or provide links to a code repository for the methodology described.
Open Datasets No The paper constructs a theoretical "simple meta-learning instance" for its analysis, rather than using a publicly available dataset with concrete access information. For example, it states: "The meta-learning instance µw is defined as uniform distribution over two tasks w and w for a fixed but unknown vector w 2 Rd."
Dataset Splits No The paper conducts theoretical analysis and does not involve empirical experiments with training, validation, or test dataset splits.
Hardware Specification No The paper does not describe any specific hardware used to run experiments or analyses.
Software Dependencies No The paper does not list any specific software dependencies with version numbers.
Experiment Setup No The paper focuses on theoretical analysis and does not describe experimental setup details such as hyperparameters or training configurations.