Enhancing Prototypical Space for Interpretable Image Classification

초록

In the Industry 4.0 era, machine learning and artificial intelligence have emerged as influential drivers in the domain of information and communication technology (ICT). This fusion of human interaction and technological progress emphasizes the necessity for artificial intelligence systems that provide transparent outcomes. The introduction of explainable artificial intelligence (XAI) plays a crucial role in establishing trust and facilitating the broader integration of AI. However, the earlier prototype-based XAI models encounter a constraint in acquiring a meaningful representation for all patches in the prototypical space. This limitation arises from its inherent update process, wherein the update gradient is only transmitted through the most active patch of the input image. To address this limitation, we introduce an innovative XAI model that introduces a novel reconstruction loss implemented by adopting a Variational autoencoder (VAE). This reconstruction loss offers insights into the rationale underlying the prototypical space and similarity mechanism within part prototypical-based XAI model, thereby enhancing comprehension of its fundamental principles. Extensive experimental results demonstrate the effectiveness of our method over existing approaches through evaluation on the widely used CUB-200-2011 dataset.

키워드