Intrinsic Explainability of Multimodal Learning for Crop Yield Prediction
Keywords: intrinsic interpretability, explainable AI, multimodal learning, Transformers, shapley values, crop yield prediction, remote sensing
TL;DR: We propose a framework for interpreting multimodal learning networks from different aspects using both intrinsic and post-hoc techniques.
Abstract: Multimodal learning enables various machine learning tasks to benefit from diverse data sources, effectively mimicking the interplay of different factors in real life events. While the heterogeneous nature of these modalities may necessitate the design of complex architectures, their interpretability is often overlooked. In this study, we leverage the intrinsic explainability of Transformer-based models to explain multimodal learning frameworks. We utilize the self-attention mechanism alongside model-specific feature attribution techniques, comparing these against post-hoc methods. Our detailed analysis focuses on the challenging task of crop yield prediction, exploiting the characteristics of the modalities and the data to aggregate local explanations at multiple levels. Our findings indicate that Transformers significantly outperform other architectures in yield prediction, making them well-suited for further intrinsic interpretability analysis. Among the modalities, satellite data emerged as the most influential but requires deeper layers for effective feature extraction due to its complex structure. Additionally, we observed that the Attention Rollout method is more robust than Generic Attention, aligns more closely with Shapley-based attributions and shows reduced sensitivity to minor input variations.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 14014
Loading