IMPACTX: Improving Model Performance by Appropriately predicting CorrecT eXplanations

TL;DR

IMPACTX addresses the challenge of using explanations to improve ML model performance in a fully automated, model-agnostic fashion. It introduces a dual-branch architecture with a top feature extractor and classifier plus a bottom latent explanation predictor and decoder that together optimize both prediction accuracy and attribution reconstruction. The approach yields consistent accuracy gains across three DL models and three image datasets (CIFAR-10/100, STL-10) and provides explanations at inference time without external XAI, with attribution maps validated via MoRF analysis. Experimental results indicate that IMPACTX can produce more reliable explanations than post-hoc methods like SHAP and ABN, suggesting a practical path toward self-explanatory, better-performing AI systems.

Abstract

The eXplainable Artificial Intelligence (XAI) research predominantly concentrates to provide explainations about AI model decisions, especially Deep Learning (DL) models. However, there is a growing interest in using XAI techniques to automatically improve the performance of the AI systems themselves. This paper proposes IMPACTX, a novel approach that leverages XAI as a fully automated attention mechanism, without requiring external knowledge or human feedback. Experimental results show that IMPACTX has improved performance respect to the standalone ML model by integrating an attention mechanism based an XAI method outputs during the model training. Furthermore, IMPACTX directly provides proper feature attribution maps for the model's decisions, without relying on external XAI methods during the inference process. Our proposal is evaluated using three widely recognized DL models (EfficientNet-B2, MobileNet, and LeNet-5) along with three standard image datasets: CIFAR-10, CIFAR-100, and STL-10. The results show that IMPACTX consistently improves the performance of all the inspected DL models across all evaluated datasets, and it directly provides appropriate explanations for its responses.

Figures (5)

• Figure 1: An overview of the IMPACTX framework. In the training phase of IMPACTX, both $M$ and $LEP$ receive $\mathbf{x}$, generating $\mathbf{m}$ and $\mathbf{z}$ respectively. These are combined for classification by $C$. In particular, $LEP$ and $D$ exploit the $R(\mathcal{A}(\mathbf{x})),\mathbf{x}, y)$ explanations. The architecture is trained using a loss function that merges MSE and CE to optimize explanation reconstruction and improve classification performance. In the inference step of IMPACTX, $C(\mathbf{m}, \mathbf{z})$ predicts the class $\hat{y}$ and $LEP-D(\mathbf{x})$ reconstructs the explanation $\mathbf{r}$ of the input $\mathbf{x}$.
• Figure 2: The Decoder architecture designed for the CIFAR-10 and CIFAR-100 datasets. The architecture is composed of convolutional (Conv2D), fully connected (FC), and UpSampling layers. The kernel size is $3 \times 3$ for all the convolutional layers, while the number of filters is given by the third dimension of the output shape.
• Figure 3: Images from the CIFAR-10 test set. The images have been filtered for better visualisation.
• Figure 4: Images from the CIFAR-100 test set. The images have been filtered for better visualisation.
• Figure 5: Images from the STL-10 test set. The images have been filtered for better visualisation.