Core Concepts
This paper proposes a method that utilizes knowledge graphs and Vector Symbolic Architecture (VSA) to enable bidirectional translation between neural network vectors and concept-level knowledge, allowing for the alignment of knowledge generated by neural networks with human-provided knowledge to enhance network training and interpretability.
Abstract
The paper addresses the challenge of leveraging the knowledge extracted from neural networks to enhance the training process. It proposes a new method that uses knowledge graphs and Vector Symbolic Architecture (VSA) to:
Convert neural network vectors into concept-level knowledge (KGVNN).
Align the knowledge graphs generated by neural networks (KGNN) with human-provided knowledge graphs (KGG) through a bipartite matching algorithm.
Use the aligned knowledge to provide feedback and supervision for optimizing the neural network.
The key aspects of the method are:
The use of knowledge graphs as the representation form to facilitate matching with human knowledge, overcoming the limitations of previous approaches that relied on ontologies or word embeddings.
The application of VSA to convert the knowledge graph matching problem into a vector matching problem, enabling efficient alignment of concepts with different names.
The introduction of auxiliary tasks and regulators to support end-to-end training and maintain the validity of the VSA-based knowledge representation.
Experiments on the MNIST dataset demonstrate the effectiveness of the proposed method in aligning neural network-generated concepts with human-provided knowledge, even when the human knowledge is incomplete or unevenly distributed. The results show that the method can consistently capture network-generated concepts that align closely with human knowledge and can even uncover new, useful concepts not previously identified by humans.
The paper highlights the potential of this approach to enhance the interpretability of neural networks and facilitate the integration of symbolic logical reasoning within these systems.