目录

• 5 ILP systems
  • 5.4 Predicate Invention
    • 5.4.2 Inverse Resolution
    • 5.4.3 Placeholders
    • 5.4.4 Metarules
    • 5.4.5 Lifelong learning
    • 5.4.6 (Unsupervised) Compression
    • 5.4.7 Connection to Representation Learning

• Oct 2020
• work in progress
• https://arxiv.org/abs/2008.07912
• A review of inductive logic programming (ILP), which is a form of machine learning dating back to the 1990s.

5 ILP systems

• Noise and Predicate Invention? ILASP, \(\partial\)ILP, Apperception.

5.4 Predicate Invention

predicate invention is attractive because it is a most natural form of automated discovery.

• example
  • grandparent, greatgrandparent. invented predicate: parent(A, B).
  • game playing, connect four
  • droplasts, higher-order, map(A, B, droplasts1)
  • ddroplasts, ddroplasts(A,B):-map(A,C,ddroplasts1),ddroplasts1(C,B).

5.4.2 Inverse Resolution

• early work

5.4.3 Placeholders

• prescripted, modeh(1,inv(person,person)). as a declaration
• Inductive Learning of Answer Set Programs (ILASP): The European Conference on Logics in Artificial Intelligence (or Journées Européennes sur la Logique en Intelligence Artificielle - JELIA) 2014
• or generates all: computationally expensive
• Learning explanatory rules from noisy data (\(\partial\)ILP): JAIR 2018
• Making sense of sensory input (Apperception): AI 2021

5.4.4 Metarules

• meta-interpretive learning, metarules (higher-order clause) to drive PI
• chain metarule, f(A,B):- tail(A,C),tail(C,B).
• unfold
• remember: no noise

5.4.5 Lifelong learning

• single-task vs. lifelong
• continually learning, dependent, easy to difficult
• reusable
• Bias reformulation for one-shot function induction: ECAI 2014
• extends to handle thousands of tasks
• Forgetting to learn logic programs (Forgetgol): AAAI 2020
• no noise
• self-supervised, Playgol, plays by randomly sampling
• Playgol: Learning programs through play: IJCAI 2019

5.4.6 (Unsupervised) Compression

• previous: measured by whether it can help solve a given task.
• does not help to solve immediately but useful
• criterion: compression
• Auto-encoding, encoder, decoder
• Program refactoring, remove redundancies
• Theory refinement: revision (correctness), compression (minimally affected), restructuring (optimize its execution or readability)

5.4.7 Connection to Representation Learning

• coincide, improving performance by changing the representation of a problem
• tabular, representing structured even relational data in such tabular form
• a few: start from the core
• Clustering-based relational unsupervised representation learning with an explicit distributed representation: IJCAI 2017
• Learning relational representations with auto-encoding logic programs: IJCAI 2019
• Lifted relational neural networks: Efficient learning of latent relational structures: JAIR 2018
• others: a propositional tabular form
• \(\partial\)ILP, neural theorem provers, short datalog programs