- Abstract: We propose Neural Logic Machines (NLMs), a neural-symbolic architecture for both inductive learning and logic reasoning. NLMs exploit the power of both neural networks—as function approximators for probabilistic distributions, and logic programming—as symbolic processor for objects with properties, relations, logic connectives, and quantifiers. After being trained on small-scale tasks (such as sorting short arrays), NLMs can learn the underlying logic rules, and generalize to arbitrarily large-scale tasks (such as sorting arbitrarily long arrays). In our experiments, NLMs achieve perfect generalization in a number of tasks, from relational reasoning tasks on family tree and general graphs, to decision making tasks including sorting, finding shortest paths, and the blocks world. Most of these tasks are hard to accomplish for neural networks or logical programming alone.
- Keywords: Neural-symbolic, first-order logic, perfect generalization
- TL;DR: A fully differentiable neural-symbolic architecture to conduct first-order logic reasoning