Scholarlink

We propose a modular deep learning framework that emphasizes compositional architectures, enabling structured generalization, efficient learning, and interpretability. By decomposing learning systems into reusable and specialized modules, our approach bridges insights from cognitive science, program synthesis, and deep learning. We demonstrate empirical improvements in sample efficiency, out-of-distribution generalization, and task transfer on several benchmarks including visual reasoning, language understanding, and reinforcement learning.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Modular Deep Learning: A Compositional Approach to Structured Generalization and Efficient Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hiep. L. Thi

摘要

We propose a modular deep learning framework that emphasizes compositional architectures, enabling structured generalization, efficient learning, and interpretability. By decomposing learning systems into reusable and specialized modules, our approach bridges insights from cognitive science, program synthesis, and deep learning. We demonstrate empirical improvements in sample efficiency, out-of-distribution generalization, and task transfer on several benchmarks including visual reasoning, language understanding, and reinforcement learning.