Automatically Composing Representation Transformations as a Means for Generalization

Abstract

A generally intelligent learner should generalize to more complex tasks than it has previously encountered, but the two common paradigms in machine learning -- either training a separate learner per task or training a single learner for all tasks -- both have difficulty with such generalization because they do not leverage the compositional structure of the task distribution. This paper introduces the compositional problem graph as a broadly applicable formalism to relate tasks of different complexity in terms of problems with shared subproblems. We propose the compositional generalization problem for measuring how readily old knowledge can be reused and hence built upon. As a first step for tackling compositional generalization, we introduce the compositional recursive learner, a domain-general framework for learning algorithmic procedures for composing representation transformations, producing a learner that reasons about what computation to execute by making analogies to previously seen problems. We show on a symbolic and a high-dimensional domain that our compositional approach can generalize to more complex problems than the learner has previously encountered, whereas baselines that are not explicitly compositional do not.

arXiv commentAccepted to the International Conference on Learning Representations 2019

Cited by in corpus (19)

• What Can Neural Networks Reason About?
• Measuring Compositional Generalization: A Comprehensive Method on Realistic Data
• Routing Networks and the Challenges of Modular and Compositional Computation
• Continual Learning in Low-rank Orthogonal Subspaces
• The Scattering Compositional Learner: Discovering Objects, Attributes, Relationships in Analogical Reasoning
• Multi-Task Reinforcement Learning with Context-based Representations
• Are Neural Nets Modular? Inspecting Functional Modularity Through Differentiable Weight Masks
• Lifelong Learning of Compositional Structures
• Dynamic Inference with Neural Interpreters
• Solving Compositional Reinforcement Learning Problems via Task Reduction
• Compositional Generalization via Neural-Symbolic Stack Machines
• Clusterability in Neural Networks
• A Study of Compositional Generalization in Neural Models
• Wide Neural Networks Forget Less Catastrophically
• Cooperative Learning of Disjoint Syntax and Semantics
• DAIR: Disentangled Attention Intrinsic Regularization for Safe and Efficient Bimanual Manipulation
• Decrypting Cryptic Crosswords: Semantically Complex Wordplay Puzzles as a Target for NLP
• Zero-shot task adaptation by homoiconic meta-mapping
• A modularity comparison of Long Short-Term Memory and Morphognosis neural networks