Modular meta-learning

Abstract

Many prediction problems, such as those that arise in the context of robotics, have a simplifying underlying structure that, if known, could accelerate learning. In this paper, we present a strategy for learning a set of neural network modules that can be combined in different ways. We train different modular structures on a set of related tasks and generalize to new tasks by composing the learned modules in new ways. By reusing modules to generalize we achieve combinatorial generalization, akin to the "infinite use of finite means" displayed in language. Finally, we show this improves performance in two robotics-related problems.

arXiv commentPresented at CoRL 2018

References in corpus (13)

• Adam: A Method for Stochastic Optimization
• Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
• Theoretical Models of Learning to Learn
• PathNet: Evolution Channels Gradient Descent in Super Neural Networks
• On First-Order Meta-Learning Algorithms
• RL${}^2$: Fast Reinforcement Learning via Slow Reinforcement Learning
• Towards a Neural Statistician
• Meta Learning Shared Hierarchies
• One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning
• Contextual Explanation Networks
• Beyond Shared Hierarchies: Deep Multitask Learning through Soft Layer Ordering
• Neural Task Programming: Learning to Generalize Across Hierarchical Tasks
• Evolutionary Architecture Search For Deep Multitask Networks

Cited by in corpus (21)

• Multi-Task Learning with Deep Neural Networks: A Survey
• A Meta-Transfer Objective for Learning to Disentangle Causal Mechanisms
• Never Stop Learning: The Effectiveness of Fine-Tuning in Robotic Reinforcement Learning
• Continual Learning in Low-rank Orthogonal Subspaces
• Neural Relational Inference with Fast Modular Meta-learning
• Meta-Learning Symmetries by Reparameterization
• Locality and compositionality in zero-shot learning
• BADGER: Learning to (Learn [Learning Algorithms] through Multi-Agent Communication)
• RoboNet: Large-Scale Multi-Robot Learning
• Experience-Embedded Visual Foresight
• Zero-Shot Learning from scratch (ZFS): leveraging local compositional representations
• S2RMs: Spatially Structured Recurrent Modules
• Decentralized Reinforcement Learning: Global Decision-Making via Local Economic Transactions
• Clusterability in Neural Networks
• Meta-Learning Dynamics Forecasting Using Task Inference
• Wide Neural Networks Forget Less Catastrophically
• Physics-aware Spatiotemporal Modules with Auxiliary Tasks for Meta-Learning
• Skeleton-based Hand-Gesture Recognition with Lightweight Graph Convolutional Networks
• A Simple Approach to Continual Learning by Transferring Skill Parameters
• Modular meta-learning in abstract graph networks for combinatorial generalization
• Learning Chebyshev Basis in Graph Convolutional Networks for Skeleton-based Action Recognition