[논문 리뷰] Multimodal Model-Agnostic Meta-Learning via Task-Aware Modulation

Risto Vuorio, Shaohua Sun외 2명|arXiv (Cornell University)|2019. 10. 30.

Domain Adaptation and Few-Shot Learning인용 수 72

한 줄 요약

MMAML은 다모달 태스크 분포에서 태스크 모드를 식별하는 조절 네트워크를 통해 MAML을 확장하고, 회귀, 분류, 강화학습에 걸친 빠른 적응을 위해 메타-학습된 priors를 조절합니다.

ABSTRACT

Model-agnostic meta-learners aim to acquire meta-learned parameters from similar tasks to adapt to novel tasks from the same distribution with few gradient updates. With the flexibility in the choice of models, those frameworks demonstrate appealing performance on a variety of domains such as few-shot image classification and reinforcement learning. However, one important limitation of such frameworks is that they seek a common initialization shared across the entire task distribution, substantially limiting the diversity of the task distributions that they are able to learn from. In this paper, we augment MAML with the capability to identify the mode of tasks sampled from a multimodal task distribution and adapt quickly through gradient updates. Specifically, we propose a multimodal MAML (MMAML) framework, which is able to modulate its meta-learned prior parameters according to the identified mode, allowing more efficient fast adaptation. We evaluate the proposed model on a diverse set of few-shot learning tasks, including regression, image classification, and reinforcement learning. The results not only demonstrate the effectiveness of our model in modulating the meta-learned prior in response to the characteristics of tasks but also show that training on a multimodal distribution can produce an improvement over unimodal training.

연구 동기 및 목표

• Identify limitation of single initialization in standard model-agnostic meta-learners under multimodal task distributions.
• Propose Multimodal Model-Agnostic Meta-Learning (MMAML) to identify task modes and modulate meta-learned priors.
• Enable rapid adaptation to new tasks via gradient updates after modulation.
• Demonstrate generalization benefits from multimodal training across regression, image classification, and reinforcement learning.

제안 방법

• A modulation network analyzes few (K) task examples to produce a task embedding v.
• Task-specific parameters tau_i are generated as tau_i = g_i(v; ω_g) for each network block i.
• Modulate each block of the task network via phi_i = θ_i ⊙ tau_i (e.g., FiLM or attention-based modulation).
• Use the modulated initialization to perform a few gradient steps to adapt to the target task, keeping tau_i fixed during adaptation.
• Train with a meta-training procedure that optimizes both the meta-learner parameters θ and the modulation network parameters (ω_h, ω_g) as in Algorithm 1.
• Domains include regression, image classification, and reinforcement learning to evaluate multimodal task adaptation.

실험 결과

연구 질문

• RQ1 Can MMAML identify task modes from few-shot task data and modulate the meta-learned prior accordingly?
• RQ2 Does training on multimodal task distributions improve generalization relative to unimodal or single-initialization meta-learners across regression, classification, and RL?
• RQ3 How does FiLM-based modulation compare to softmax-based attention for modulating network parameters?
• RQ4 What is the impact of MMAML on fast adaptation performance versus MAML and Multi-MAML across different modes or datasets?

주요 결과

• MMAML with FiLM modulation outperforms standard MAML and achieves competitive results with Multi-MAML across multimodal regression and image classification benchmarks.
• Task embeddings learned by the modulation network cluster according to task modes, indicating successful mode identification.
• FiLM modulation yields more stable and superior performance than softmax attention in the experimented domains.
• MMAML benefits from multimodal training, showing better generalization than unimodal-trained baselines in several tasks.
• In reinforcement learning, MMAML consistently surpasses unmodulated ProMP baselines, with performance improving as task modes increase.