Four Challenges in the Blocks World Problem
- Recover a set of lifted rules and generalize to worlds with more blocks;
- Deal with high-order relational data and quantifiers;
- Scale up w.r.t. the complexity of the rules;
- Complexity: O(m^B * D * C^2)
- Parameters: O(D * C^2)
- Recover rules based on a minimal set of learning priors.
Neural Logic Machines
- Logic Predicates as Tensors
- Logic Rules as Neural Operators
- Boolean Logic Rules: AND, OR, NOT
- Quantifications: \any, \exists
Curriculum Learning Guided by Exams and Fails
Pros & Cons
- With polynomial complexity and parameter size NLMs can deal with more complex task which requires using a large number of intermediate predicates.
- NLMs are general neural architectures for learning lifted rules from only input-output pairs without pre-designed rules nor knowledge base.
- The scalability of NLMs to large entity sets in systems of KB reasoning is left as future work.
- NLMs can be plugged into existing convolutional or recurrent neural architectures for logic reasoning.
- Adaptively select the right maximum depth for novel problems;
- Handle vector inputs, not only symbolic inputs;
- Optimize the training to be effective and simple;
- Extract human-readable rules from weights.
Page Not Found
Try to search through the entire repo.