Interpretable Tensor Fusion

Saurabh Varshneya, Antoine Ledent, Philipp Liznerski, Andriy Balinskyy, Purvanshi Mehta, Waleed Mustafa, Marius Kloft

Abstract

Conventional machine learning methods are predominantly designed to predict outcomes based on a single data type. However, practical applications may encompass data of diverse types, such as text, images, and audio. We introduce interpretable tensor fusion (InTense), a multimodal learning method training a neural network to simultaneously learn multiple data representations and their interpretable fusion. InTense can separately capture both linear combinations and multiplicative interactions of the data types, thereby disentangling higher-order interactions from the individual effects of each modality. InTense provides interpretability out of the box by assigning relevance scores to modalities and their associations, respectively. The approach is theoretically grounded and yields meaningful relevance scores on multiple synthetic and real-world datasets. Experiments on four real-world datasets show that InTense outperforms existing state-of-the-art multimodal interpretable approaches in terms of accuracy and interpretability.

Introduction

In the real world, data rarely comes in a single format. To detect complex social signals like sarcasm, a machine must integrate visual cues (a smirk), auditory signals (an animated tone), and textual content (a sentence that may appear positive but is biting in context). While multimodal learning has advanced significantly, many state-of-the-art fusion techniques remain "black boxes." We know they work, but we often don't know why—specifically, which combinations of text, audio, or video drove the decision.

Conceptual overview of multimodal fusion for sarcasm detection, distinguishing between linear and non-linear interactions.

A major challenge in multimodal fusion is higher-order interaction bias: the tendency for models to overstate the importance of complex interactions even when a simple linear combination of modalities would suffice. This leads to spurious interpretations where a model appears to be performing sophisticated fusion while actually overfitting to redundant signals.

Methodology: Disentangling Interactions

We introduce InTense (Interpretable Tensor Fusion), which addresses these challenges through two key pillars:

  1. Multiple Neural Learning (MNL): This deep-learning adaptation of classic Multiple Kernel Learning allows the network to learn distinct representations for each modality. A fusion layer then assigns a weight, or relevance score ($\beta$), to each modality and their possible interactions.
  2. Iterative Batch Normalization (IterBN): To solve the interaction bias, we developed IterBN. This normalization scheme ensures that higher-order interactions (e.g., Video $\otimes$ Audio $\otimes$ Text) cannot "cheat" by capturing information already present in individual modalities. It forces the model to only assign high relevance to an interaction if that interaction provides unique, non-linear value.

The resulting relevance scores are mathematically derived from the weights of the fusion layer using $p$-norms, making interpretability an inherent property of the model rather than a post-hoc approximation.

Experiments and Results

InTense was evaluated on a diverse set of benchmarks, including MUStARD (sarcasm detection), CMU-MOSI and CMU-MOSEI (sentiment analysis), UR-FUNNY (humor detection), AV-MNIST, and ENRICO.

Superior Accuracy and Interpretability

InTense outperformed existing interpretable baselines, such as MultiRoute and MRO, across all six real-world datasets. Furthermore, it performed within 2% of the accuracy of the Multimodal Transformer (MulT), a leading "black box" model, suggesting that transparency does not require a significant sacrifice in performance.

Relevance scores across four real-world datasets, showing the importance of interactions in tasks like sarcasm detection.

The model's relevance scores provide actionable insights into the nature of different tasks. For example, in sentiment analysis (CMU-MOSI), text is typically the dominant modality. However, in sarcasm detection (MUStARD), InTense revealed that interactions between modalities (the "vat" score) are significantly more important than any single modality alone. This confirms the intuition that sarcasm is a holistic phenomenon requiring the "melt" of multiple signals.

Conclusion

InTense provides a theoretically grounded approach to multimodal learning that offers both high accuracy and "out-of-the-box" interpretability. By disentangling individual effects from complex interactions, it allows researchers to peek inside the multimodal black box. This makes InTense particularly well-suited for safety-critical domains like medicine or finance, where understanding the drivers of a model's decision is mandatory.

Summary updated

Abstract

Conventional machine learning methods are predominantly designed to predict outcomes based on a single data type. However, practical applications may encompass data of diverse types, such as text, images, and audio. We introduce interpretable tensor fusion (InTense), a multimodal learning method training a neural network to simultaneously learn multiple data representations and their interpretable fusion. InTense can separately capture both linear combinations and multiplicative interactions of the data types, thereby disentangling higher-order interactions from the individual effects of each modality. InTense provides interpretability out of the box by assigning relevance scores to modalities and their associations, respectively. The approach is theoretically grounded and yields meaningful relevance scores on multiple synthetic and real-world datasets. Experiments on four real-world datasets show that InTense outperforms existing state-of-the-art multimodal interpretable approaches in terms of accuracy and interpretability.
BibTeX
@inproceedings{varshneya2024interpretable,
  author = {Saurabh Varshneya and Antoine Ledent and Philipp Liznerski and Andriy Balinskyy and Purvanshi Mehta and Waleed Mustafa and Marius Kloft},
  title = {{Interpretable Tensor Fusion}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=sf8LBVUb64},
}
Figure