Semantic-based neural network repair for AI models in dynamic environments

Abstract

This paper addresses a critical challenge in the field of artificial intelligence (AI) and deep learning (DL) - the debugging and repair of AI models within dynamic environments. The rapid development of frameworks like TensorFlow or PyTorch can introduce version incompatibilities, potentially breaking existing code. To tackle this issue, it can be seen that a semantic-based neural network repair method that leverages an existing semantics called ExAIS is proposed, defined in the Prolog logic programming language. This method identifies architectural and parameter errors in AI models and generates debug messages to facilitate the repair process. Key contributions of this paper include introducing a novel semantic-based AI model repair method capable of suggesting multiple changes to rectify inefficiencies in AI models. We evaluate the effectiveness of our approach in a real-world context by collecting a set of real bugs from AI developers. The background section provides an overview of AI framework testing, machine learning, neural architecture search (NAS), and deep learning. The methodology section details the repair method based on ExAIS semantics, covering error identification and repair suggestions as the two main steps. The evaluation section showcases the repair effectiveness on automatically generated and manually written models, emphasizing the applicability and efficiency of the approach.