M²AML: Metric-Based Model-Agnostic Meta-Learning for Few-Shot Classification

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Model-Agnostic Meta-Learning (MAML) and Prototypical Networks (ProtoNet) are key few-shot classification methods.
• MAML faces optimization instability; ProtoNet lacks task-specific adaptation.
• Existing methods struggle with domain shifts and adaptation efficiency.

Purpose of the Study:

• Introduce Metric-based Model-Agnostic Meta-Learning (M²AML) to address limitations of MAML and ProtoNet.
• Enhance few-shot classification performance and adaptation speed.
• Provide a stable and efficient meta-learning framework.

Main Methods:

• Developed M²AML, removing parameterized classification layers from episodic adaptation.
• Replaced inner-loop classification with a dynamic self-exclusive geometric similarity metric.
• Optimized spatial distances instead of functional mappings for efficient adaptation.

Main Results:

• M²AML demonstrated state-of-the-art performance across mini-ImageNet, tiered-ImageNet, and CIFAR-FS datasets.
• Achieved absolute accuracy improvements of 0.1% to 2.1% over leading models.
• Showcased synchronized inner/outer learning rates and accelerated adaptation steps.

Conclusions:

• M²AML effectively reconciles structural limitations of prior meta-learning algorithms.
• The geometric similarity metric offers a robust alternative for few-shot classification.
• M²AML presents a significant advancement in meta-learning for computer vision tasks.