Connectionist Knowledge Representation and Reasoning

1
Connectionist Knowledge Representation and
Reasoning
SCREECH

• Barbara Hammer
• Computer Science, Clausthal University of
  Technology, Germany
• Pascal Hitzler
• AIFB, Universiy of Karlsruhe, Germany

2
General Motivation
connectionist
knowledge representation and reasoning

• Artificial Neural Networks and Symbolic Knowledge
  Representation and Reasoning are two diverse
  Paradigms in Artificial Intelligence.
• Their strengths and weaknesses complement each
  other.
• We seek to combine them in order to obtain
  systems with functionalities being the best of
  both worlds.

3
Artificial Neural Networks (ANNs)

• Powerful machine learning paradigm.
• Architectures inspired by Biology.
• Can be trained on raw and noisy data.
• Robust. Graceful degradation.
• No declarative reading. Black boxes.
• Dealing with recursive structures difficult.
• Training cannot take a priori domain knowledge
  into account.

?
?
4
Knowledge Representation and Reasoning (KRR)

• Logic-based. Declarative.
• Modelling inspired by human thinking.
• Simple manual coding of knowledge.
• Highly recursive.
• Systems hard to train.
• No tolerance to noise. Brittle.
• Reasoning algorithms with high complexities.

?
?
5
connectionist
knowledge representation and reasoning
?
?
?
?
6
Issues in Connectionist KRR

• Representation of symbolic knowledge within ANNs.
• Extraction of symbolic knowledge from ANNs.
• Learning of symbolic knowledge using ANNs.
• Learning taking symbolic background knowledge
  into account.

7
Tutorial Outline

• Part I Neural networks and structured knowledge
• Feedforward networks
• Recurrent networks
• Recursive data structures
• Part II Logic and neural networks
• Propositional logic
• First-order logic
• Future challenges