[論文レビュー] LGM-Net: Learning to Generate Matching Networks for Few-Shot Learning

Huaiyu Li, Weiming Dong+4 more|arXiv (Cornell University)|May 15, 2019

Domain Adaptation and Few-Shot Learning被引用数 50

ひとこと要約

LGM-Net は MetaNet を訓練して few-shot タスクデータから TaskNet 重みを生成し、ファインチューニングなしで未見タスクへ迅速に適応できるようにする。Task Context Encoder と Weight Generator を用いて TargetNet パラメータを生成し、intertask normalization によりタスク間の情報を共有。

ABSTRACT

In this work, we propose a novel meta-learning approach for few-shot classification, which learns transferable prior knowledge across tasks and directly produces network parameters for similar unseen tasks with training samples. Our approach, called LGM-Net, includes two key modules, namely, TargetNet and MetaNet. The TargetNet module is a neural network for solving a specific task and the MetaNet module aims at learning to generate functional weights for TargetNet by observing training samples. We also present an intertask normalization strategy for the training process to leverage common information shared across different tasks. The experimental results on Omniglot and miniImageNet datasets demonstrate that LGM-Net can effectively adapt to similar unseen tasks and achieve competitive performance, and the results on synthetic datasets show that transferable prior knowledge is learned by the MetaNet module via mapping training data to functional weights. LGM-Net enables fast learning and adaptation since no further tuning steps are required compared to other meta-learning approaches.

研究の動機と目的

• Motivate and address the need for rapid adaptation to unseen few-shot tasks by leveraging transferable prior knowledge across tasks.
• Propose a meta-learning framework that directly generates functional weights for a TaskNet from limited task data.
• Introduce an efficient MetaNet architecture with a Task Context Encoder and a conditional Weight Generator.
• Incorporate intertask normalization to exploit shared information across tasks during training.

提案手法

• Two-module architecture: TargetNet (the TaskNet using generated weights) and MetaNet (generates TargetNet weights from task data).
• MetaNet comprises a Task Context Encoder that encodes training samples into a fixed-size context and a Conditional Weight Generator that maps the context to TargetNet weights.
• Task context is modeled as a reparameterized multivariate Gaussian; weights for each TargetNet layer are produced by layer-specific generators and normalized (weight normalization).
• TargetNet is a Matching Network whose parameters are generated by MetaNet; classification uses cosine-distance based attention (attentional metric) over embedded features.
• Intertask Normalization (ITN) via batch normalization across a batch of tasks to share statistics and improve training.
• Training procedure optimizes MetaNet across episodic tasks using the meta-training data with cross-entropy loss on the test splits of each task.

実験結果

リサーチクエスチョン

• RQ1 Can a meta-learner learn to generate the functional weights of a TaskNet from limited task data and generalize to unseen tasks?
• RQ2 Does generating TaskNet weights via MetaNet improve few-shot performance compared to traditional meta-learning methods that rely on initialization or update rules?
• RQ3 What is the impact of the Task Context Encoder and Intertask Normalization on generalization to new tasks?
• RQ4 How do generated weights distribute across tasks and what does this imply about transferable prior knowledge?

主な発見

Model	5-way 1-shot	5-way 5-shot	20-way 1-shot
Matching networks (Vinyals et al., 2016)	43.56 ± 0.84%	55.31 ± 0.73%	17.31 ± 0.22%
Meta-LSTM (Ravi & Larochelle, 2017)	43.44 ± 0.77%	60.60 ± 0.71%	16.70 ± 0.23%
MetaNet (Munkhdalai & Yu, 2017)	49.21 ± 0.96%	-	-
Prototypical Nets (Snell et al., 2017)	49.42 ± 0.78%	68.20 ± 0.66%	-
MAML (Finn et al., 2017)	48.70 ± 1.84%	63.11 ± 0.92%	16.49 ± 0.58%
Meta-SGD (Li et al., 2017)	50.47 ± 1.87%	64.03 ± 0.94%	17.56 ± 0.64%
Relation Net (Sung et al., 2018)	51.38 ± 0.82%	67.07 ± 0.69%	-
REPTILE (Nichol & Schulman)	49.97 ± 0.32%	65.99 ± 0.58%	-
SNAIL (Mishra et al., 2018)	55.71 ± 0.99%	65.99 ± 0.58%	-
(Gidaris & Komodakis, 2018)	56.20 ± 0.86%	73.00 ± 0.64%	-
LEO (Rusu et al., 2019)	61.76 ± 0.08%	77.59 ± 0.12%	-
LGM-Net (Ours)	69.13 ± 0.35%	71.18 ± 0.68%	26.14 ± 0.34%

• On mini-ImageNet, LGM-Net achieves state-of-the-art-like results, with 5-way 1-shot at 69.13% and 5-way 5-shot at 71.18%, surpassing several baselines.
• On Omniglot, LGM-Net achieves competitive performance across 5-way and 20-way settings (e.g., 99.0% at 5-way 1-shot).
• Ablation shows ITN significantly improves performance, Task Context Encoder contributes beyond random priors, and weight normalization stabilizes training.
• Generated weights cluster by task in t-SNE visualizations, indicating the MetaNet learns task-specific weight distributions and transfers to similar tasks.
• Compared to fixed-weight matching networks and several meta-learning baselines, LGM-Net demonstrates improved adaptability and faster inference since no extra fine-tuning is required.