Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Step towards Interpretable Multimodal AI Models with MultiFIX

Mafalda Malafaia, Thalea Schlender, Tanja Alderliesten, Peter A. N. Bosman

TL;DR

This work tackles the challenge of building interpretable multimodal AI by introducing MultiFIX, a pipeline that uses DL to learn features from heterogeneous data and GP-GOMEA to evolve symbolic expressions for tabular features and final fusion. It supports Grad-CAM explanations for image blocks and yields intrinsically interpretable models for tabular data and fusion, aiming to preserve predictive accuracy. Through experiments on synthetic image+tabular problems with varying modality interactions, MultiFIX consistently outperforms single-modality baselines and provides faithful explanations that align with the ground-truth features, with end-to-end and hybrid training often delivering the best performance. The results demonstrate the feasibility of interpretable multimodal learning and offer guidance on training strategies, while also outlining limitations and avenues for future work, including real-world validation and richer, inherently interpretable image blocks.

Abstract

Real-world problems are often dependent on multiple data modalities, making multimodal fusion essential for leveraging diverse information sources. In high-stakes domains, such as in healthcare, understanding how each modality contributes to the prediction is critical to ensure trustworthy and interpretable AI models. We present MultiFIX, an interpretability-driven multimodal data fusion pipeline that explicitly engineers distinct features from different modalities and combines them to make the final prediction. Initially, only deep learning components are used to train a model from data. The black-box (deep learning) components are subsequently either explained using post-hoc methods such as Grad-CAM for images or fully replaced by interpretable blocks, namely symbolic expressions for tabular data, resulting in an explainable model. We study the use of MultiFIX using several training strategies for feature extraction and predictive modeling. Besides highlighting strengths and weaknesses of MultiFIX, experiments on a variety of synthetic datasets with varying degrees of interaction between modalities demonstrate that MultiFIX can generate multimodal models that can be used to accurately explain both the extracted features and their integration without compromising predictive performance.

A Step towards Interpretable Multimodal AI Models with MultiFIX

TL;DR

This work tackles the challenge of building interpretable multimodal AI by introducing MultiFIX, a pipeline that uses DL to learn features from heterogeneous data and GP-GOMEA to evolve symbolic expressions for tabular features and final fusion. It supports Grad-CAM explanations for image blocks and yields intrinsically interpretable models for tabular data and fusion, aiming to preserve predictive accuracy. Through experiments on synthetic image+tabular problems with varying modality interactions, MultiFIX consistently outperforms single-modality baselines and provides faithful explanations that align with the ground-truth features, with end-to-end and hybrid training often delivering the best performance. The results demonstrate the feasibility of interpretable multimodal learning and offer guidance on training strategies, while also outlining limitations and avenues for future work, including real-world validation and richer, inherently interpretable image blocks.

Abstract

Real-world problems are often dependent on multiple data modalities, making multimodal fusion essential for leveraging diverse information sources. In high-stakes domains, such as in healthcare, understanding how each modality contributes to the prediction is critical to ensure trustworthy and interpretable AI models. We present MultiFIX, an interpretability-driven multimodal data fusion pipeline that explicitly engineers distinct features from different modalities and combines them to make the final prediction. Initially, only deep learning components are used to train a model from data. The black-box (deep learning) components are subsequently either explained using post-hoc methods such as Grad-CAM for images or fully replaced by interpretable blocks, namely symbolic expressions for tabular data, resulting in an explainable model. We study the use of MultiFIX using several training strategies for feature extraction and predictive modeling. Besides highlighting strengths and weaknesses of MultiFIX, experiments on a variety of synthetic datasets with varying degrees of interaction between modalities demonstrate that MultiFIX can generate multimodal models that can be used to accurately explain both the extracted features and their integration without compromising predictive performance.
Paper Structure (29 sections, 1 equation, 4 figures, 7 tables)

This paper contains 29 sections, 1 equation, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. MultiFIX: Multimodal Feature engIneering for eXplainable AI
  5. Feature Engineering Blocks
  6. Fusion Strategy
  7. Training Strategies
  8. Interpretability Techniques
  9. Experimental Setup
  10. Dataset
  11. Experiments and Results
  12. AND Problem
  13. Problem Description
  14. DL Performance Results
  15. Interpretable Model
  16. ...and 14 more sections

Figures (4)

  • Figure 1: Overview of MultiFIX. Data passes into the feature engineering blocks. Feature vectors I and T are concatenated and passed to the fusion block to make the final prediction in the Training Stage (top). In the Inference Stage, image features are explained through Grad-CAM, and symbolic expressions are obtained for both the tabular features and the target prediction with GP-GOMEA, replacing their NN counterparts.
  • Figure 2: Representative samples for the image modality.
  • Figure 3: Interpretable Models: Grad-CAM heatmaps explain the image input contributions for each extracted feature. GP-GOMEA symbolic expressions explain the tabular features and the fusion of both modalities to make the prediction. Learned features and predictions are compared to their Ground Truth (GT) counterparts.
  • Figure 4: Interpretable Models: Grad-CAM heatmaps explain the image input contributions for each extracted feature. GP-GOMEA symbolic expressions explain the tabular features and the fusion of both modalities to make the prediction. Learned features and predictions are compared to their GT counterparts.