[Paper Review] Meta-Learning with Warped Gradient Descent

Sebastian Flennerhag, Andrei A. Rusu+4 more|arXiv (Cornell University)|Aug 30, 2019

Domain Adaptation and Few-Shot Learning64 references66 citations

TL;DR

WarpGrad meta-learns warp-layers to precondition gradients, enabling scalable, trajectory-agnostic gradient-based meta-learning across few-shot, standard, continual, and reinforcement learning tasks.

ABSTRACT

Learning an efficient update rule from data that promotes rapid learning of new tasks from the same distribution remains an open problem in meta-learning. Typically, previous works have approached this issue either by attempting to train a neural network that directly produces updates or by attempting to learn better initialisations or scaling factors for a gradient-based update rule. Both of these approaches pose challenges. On one hand, directly producing an update forgoes a useful inductive bias and can easily lead to non-converging behaviour. On the other hand, approaches that try to control a gradient-based update rule typically resort to computing gradients through the learning process to obtain their meta-gradients, leading to methods that can not scale beyond few-shot task adaptation. In this work, we propose Warped Gradient Descent (WarpGrad), a method that intersects these approaches to mitigate their limitations. WarpGrad meta-learns an efficiently parameterised preconditioning matrix that facilitates gradient descent across the task distribution. Preconditioning arises by interleaving non-linear layers, referred to as warp-layers, between the layers of a task-learner. Warp-layers are meta-learned without backpropagating through the task training process in a manner similar to methods that learn to directly produce updates. WarpGrad is computationally efficient, easy to implement, and can scale to arbitrarily large meta-learning problems. We provide a geometrical interpretation of the approach and evaluate its effectiveness in a variety of settings, including few-shot, standard supervised, continual and reinforcement learning.

Motivation & Objective

• Motivate and address limitations of existing gradient-based meta-learners (convergence, scalability, and credit assignment).
• Propose a trajectory-agnostic preconditioning framework that embeds warp-layers between task-learner layers to precondition gradients.
• Provide a geometrical interpretation of WarpGrad via Riemannian metrics and demonstrate scalable performance across few-shot, multi-shot, continual, and reinforcement learning settings.

Proposed method

• Introduce warp-layers interleaved with task-learner layers to form a warped network, enabling data-dependent preconditioning of gradients.
• Define a general preconditioning rule U(θ;φ)=θ−αP(θ;φ)∇L(θ), where P is realized via warp-layers and their Jacobians.
• Derive a trajectory-agnostic meta-objective L(φ) by optimizing over a joint distribution of tasks and intermediate task-learner parameters, avoiding backpropagation through the full adaptation trajectory.
• Explain the geometry: warp-layers induce a metric G on the warping space, with G−1 acting as the preconditioner; establish first-order equivalence between updates in the warped space and descent under the Riemannian metric.
• Propose online (Algorithm 1) and offline (Algorithm 2) meta-training procedures to learn warp-parameters φ while optionally learning or using a prior over initial task parameters θ0|τ.
• Show integration with learned initialisations and priors, enabling various training regimes (online/offline, supervised/RL continual learning).
• Demonstrate non-linear warp-layers to capture richer preconditioning beyond block-diagonal structures and exhibit memory-capable behavior in RL tasks.

Experimental results

Research questions

• RQ1 Can WarpGrad retain the inductive bias of gradient-based few-shot learners while avoiding backpropagation through adaptation trajectories?
• RQ2 To what extent can WarpGrad scale beyond few-shot learning to multi-shot and standard supervised/ RL tasks?
• RQ3 DoesWarpGrad generalize to complex meta-learning scenarios such as continual learning and tasks requiring memory?
• RQ4 Is the learned warp-geometry interpretable as a curvature-based preconditioner facilitating convergence guarantees?

Key findings

• WarpGrad outperforms baseline gradient-based meta-learners on standard few-shot benchmarks (mini-ImageNet and tiered-ImageNet).
• Warp-MAML and Warp-Leap variants achieve higher accuracy than their non-warp counterparts in few-shot and multi-shot settings, including Omniglot and tiered-ImageNet with extended adaptation steps.
• Non-linear warp-layers enable richer preconditioning beyond block-diagonal structures, improving performance on complex tasks such as continual learning and RL maze navigation.
• Offline meta-training with warp-learned parameters yields substantial gains (e.g., Omniglot results improving from 76.3% to 84.3% test accuracy).
• WarpGrad maintains convergence properties by embedding preconditioning as a gradient-descent-like update via an implicit Riemannian metric, providing stability and scalability across tasks.