[Paper Review] Fast Context Adaptation via Meta-Learning

Luisa Zintgraf, Kyriacos Shiarlis+3 more|arXiv (Cornell University)|Oct 8, 2018

Machine Learning and Data Classification176 citations

TL;DR

CAVIA extends MAML by splitting model parameters into task-specific context parameters and shared meta-learned parameters, updating only the context at test time to achieve fast, robust adaptation with interpretable task embeddings across regression, classification, and reinforcement learning.

ABSTRACT

We propose CAVIA for meta-learning, a simple extension to MAML that is less prone to meta-overfitting, easier to parallelise, and more interpretable. CAVIA partitions the model parameters into two parts: context parameters that serve as additional input to the model and are adapted on individual tasks, and shared parameters that are meta-trained and shared across tasks. At test time, only the context parameters are updated, leading to a low-dimensional task representation. We show empirically that CAVIA outperforms MAML for regression, classification, and reinforcement learning. Our experiments also highlight weaknesses in current benchmarks, in that the amount of adaptation needed in some cases is small.

Motivation & Objective

• Advance fast adaptation in meta-learning with reduced meta-overfitting and improved parallelism.
• Introduce context parameters that adapt per task while sharing other parameters across tasks.
• Demonstrate that context parameters yield a low-dimensional, interpretable task embedding.
• Show that CAVIA outperforms MAML on regression, classification, and reinforcement learning tasks.

Proposed method

• Partition model parameters into context parameters phi and shared parameters theta.
• Inner loop updates adjust only phi for each task, using a gradient step on the task loss.
• Outer loop updates adjust theta by backpropagating through the inner-loop phi update.
• Context parameters are initialised to zero and conditioned into the network (e.g., via input augmentation or FiLM).
• Meta-learning optimizes theta to perform well after the inner-loop adaptation across tasks.
• At test time, only phi is updated to adapt to a new task while theta remains fixed.

Experimental results

Research questions

• RQ1 Can updating only a small context vector per task match or exceed MAML performance across domains?
• RQ2 Do context parameters learn meaningful, task-specific embeddings that capture latent structure across tasks?
• RQ3 Is CAVIA robust to a range of inner-loop learning rates and scalable to larger networks without overfitting?
• RQ4 How does CAVIA’s parallelization and memory usage compare to MAML in practice?

Key findings

• CAVIA outperforms MAML on regression, classification, and reinforcement learning benchmarks.
• Adapting only a small number of context parameters (2–50) suffices; larger networks can be used without inner-loop overfitting.
• Context parameters learn meaningful embeddings of tasks, enabling interpolation between tasks.
• CAVIA is robust to inner-loop learning rate choices and scales better than MAML when increasing network size.