Few-Shot Induction of Generalized Logical Concepts via Human Guidance

Area of Science:

• Artificial Intelligence
• Machine Learning
• Logic Programming

Background:

• Learning generalized first-order representations from limited data is challenging.
• Inductive logic programming (ILP) is a key technique for symbolic representation learning.

Purpose of the Study:

• To improve the efficiency and sample efficiency of ILP for learning generalized first-order concept representations.
• To introduce novel methods for concept representation and human-in-the-loop learning within ILP.

Main Methods:

• Augmenting an ILP learner with a novel distance measure between candidate concept representations.
• Incorporating human advice as richer input to guide the learning process.
• Proving the semantic validity of the distance measure and deriving a Probably Approximately Correct (PAC) bound.

Main Results:

• The proposed distance measure significantly enhances search efficiency for target concepts and generalizations.
• Leveraging human advice improves the sample efficiency of the learning process.
• Experiments on diverse tasks validate the effectiveness and efficiency of the developed approach.

Conclusions:

• The novel distance measure and human advice integration offer a robust solution for few-shot learning of generalized first-order representations.
• The approach demonstrates practical applicability and theoretical soundness through semantic validation and PAC bounds.
• This work advances the capabilities of ILP in handling complex concept learning scenarios with limited data.