Towards a Unifying Logical Framework for Neural Networks

Xiyue Zhang; Xiaohong Chen; Meng Sun

Permalink

https://hdl.handle.net/2142/114412 Copy

Description

Title

Towards a Unifying Logical Framework for Neural Networks

Author(s)

• Xiyue Zhang
• Xiaohong Chen
• Meng Sun

Issue Date

2022-08-15

Keyword(s)

• Matching logic
• Neural networks
• Formal specifications

Abstract

Neural networks are increasingly used in safety-critical applications such as medical diagnosis and autonomous driving, which calls for the need for formal specification of their behaviors. In this paper, we use matching logic—a unifying logic to specify and reason about programs and computing systems—to axiomatically define dynamic propagation and temporal operations in neural networks and to formally specify common properties about neural networks. As instances, we use matching logic to formalize a variety of neural networks, including generic feedforward neural networks with different activation functions and recurrent neural networks. We define their formal semantics and several common properties in matching logic. This way, we obtain a unifying logical framework for specifying neural networks and their properties.

Type of Resource

text

Language

en

Handle URL

https://hdl.handle.net/2142/114412

Permalink

https://hdl.handle.net/2142/114412

Owning Collections