Neural Networks as Artificial Specifications

TL;DR

This paper explores training neural networks as artificial specifications for programs, addressing previous challenges like false positives and training difficulties by examining various training modes and abstractions, with promising results.

Contribution

It investigates unexamined factors affecting neural network training for program specifications, improving understanding of their practical viability.

Findings

Different learning modes impact training success

Abstraction functions influence false positive rates

Training approaches show promising results

Abstract

In theory, a neural network can be trained to act as an artificial specification for a program by showing it samples of the programs executions. In practice, the training turns out to be very hard. Programs often operate on discrete domains for which patterns are difficult to discern. Earlier experiments reported too much false positives. This paper revisits an experiment by Vanmali et al. by investigating several aspects that were uninvestigated in the original work: the impact of using different learning modes, aggressiveness levels, and abstraction functions. The results are quite promising.