Improving AI Deployment through Interpretability and Explainability Algorithms

The significant evolution that systems based on Artificial Intelligence has experienced in the last decade has been followed by research on the analysis of the internals of these models. More specifically, what characteristics or features of the training data are internally encoded in the model (Interpretability); and, what characteristics from a given input are considered by the model when making a prediction for such input (Explainability).

As part of the ACRAI Research Day 2025, we will contribute with two talks presenting our research on how these Interpretability and Explainability methods can be transferred to assist more practical settings where AI-based systems are deployed.

Talk-1: Deep Model Interpretation with Limited Data (Hamed Behzadi-Khormouji, José Oramas)

In the first talk, we present an approach to exploit "dataset summaries", produced via Coreset Selection Algorithms, as a means to reduce the computational costs of model interpretation algorithms. Thus, allowing the frequent application of these interpretation methods during the development and training of a model; enabling their use as a tool for debugging.

Talk-2: Explainability-based Band Selection for Hyperspectral Image Analysis (Salma Haidar, José Oramas)

In the second talk, in the context of Hyperspectral Image Analysis (HSI), we present how insights extracted from Explainability Algorithms can help the selection of most relevant information (bands) from the input. Thus, enabling the development of HSI systems with shorter image acquisition and processing times and lower memory requirements.