ExplainabilityTransparencyAccountabilitycs.LG
Towards a general framework for improving the performance of classifiers using XAI methods
Andrea Apicella, Salvatore Giugliano, Francesco Isgrò, Roberto Prevete
March 15, 2024
arXiv.org
Key Takeaways
-
XAI can serve a dual purpose beyond transparency: While explainable AI is typically used to help humans understand AI decisions, this research demonstrates that XAI techniques can also be repurposed to automatically improve an AI system's performance—offering a practical benefit beyond accountability.
-
Performance improvements without expensive retraining: The framework allows organizations to enhance existing AI classifiers without the significant computational costs of retraining models from scratch, making AI improvement more accessible and sustainable for resource-constrained organizations.
-
Two complementary technical approaches exist: The paper identifies auto-encoder-based and encoder-decoder-based strategies as viable methods for this improvement process, suggesting that different organizations may find different approaches suitable depending on their specific needs and constraints.
-
Bridges the gap between interpretability and utility: This work addresses a growing research interest in making XAI methods not just explanatory tools for compliance and trust, but also functional tools that directly enhance AI system reliability and accuracy—potentially aligning business incentives with responsible AI practices.
Abstract
Modern Artificial Intelligence (AI) systems, especially Deep Learning (DL) models, poses challenges in understanding their inner workings by AI researchers. eXplainable Artificial Intelligence (XAI) inspects internal mechanisms of AI models providing explanations about their decisions. While current XAI research predominantly concentrates on explaining AI systems, there is a growing interest in using XAI techniques to automatically improve the performance of AI systems themselves. This paper proposes a general framework for automatically improving the performance of pre-trained DL classifiers using XAI methods, avoiding the computational overhead associated with retraining complex models from scratch. In particular, we outline the possibility of two different learning strategies for implementing this architecture, which we will call auto-encoder-based and encoder-decoder-based, and discuss their key aspects.