Meta-Learning With Differentiable Convex Optimization

TLDR

Many meta‑learning methods for few‑shot learning use simple base learners such as nearest‑neighbor classifiers, yet discriminatively trained linear predictors can generalize better even with few shots. The authors aim to learn feature embeddings that generalize well under a linear classification rule for novel categories and to use these discriminatively trained linear predictors as base learners, demonstrating improved tradeoffs between feature size and performance on few‑shot recognition benchmarks. They solve the objective by exploiting implicit differentiation of the convex optimality conditions and the dual formulation of linear classifiers. The resulting MetaOptNet uses high‑dimensional embeddings with improved generalization and modest computational overhead, achieving state‑of‑the‑art performance on miniImageNet, tieredImageNet, CIFAR‑FS, and FC100 few‑shot learning benchmarks.

Abstract

Many meta-learning approaches for few-shot learning rely on simple base learners such as nearest-neighbor classifiers. However, even in the few-shot regime, discriminatively trained linear predictors can offer better generalization. We propose to use these predictors as base learners to learn representations for few-shot learning and show they offer better tradeoffs between feature size and performance across a range of few-shot recognition benchmarks. Our objective is to learn feature embeddings that generalize well under a linear classification rule for novel categories. To efficiently solve the objective, we exploit two properties of linear classifiers: implicit differentiation of the optimality conditions of the convex problem and the dual formulation of the optimization problem. This allows us to use high-dimensional embeddings with improved generalization at a modest increase in computational overhead. Our approach, named MetaOptNet, achieves state-of-the-art performance on miniImageNet, tieredImageNet, CIFAR-FS, and FC100 few-shot learning benchmarks.