Learning Reliable Rules by Re-generating Deep Features
Keywords: Interpretable ML, Neuro-symbolic AI, SATNet, Logical Reasoning
Abstract: Improving the interpretability and reliability of deep learning models is essential for advancing machine learning applications, though it remains a significant challenge. One promising approach is the integration of logical reasoning into deep learning systems. Previous works have demonstrated that SATNet, a differentiable MaxSAT solver, can learn interpretable and reliable rules from input-output examples in puzzle domains. In this work, we propose *Visual SATNet* (Vi-SATNet), an extended version of SATNet capable of learning logical reasoning rules in more general and complex domains, such as the feature space of real-life images. We find that, given a pre-trained deep convolutional neural network (CNN) architecture, a Vi-SATNet layer can be integrated and trained efficiently to learn a set of reasoning rules on the deep features, guiding the classifier’s decision. Vi-SATNets are trained to perform feature re-generation tasks for a given image dataset, where the re-generated features maintain high accuracy when used for image classification, proving their quality. In our experiment on the Imagenette dataset with a pre-trained VGG19 model, masking out 10\% to 80\% of the features results in classification accuracy ranging from 98.50\% to 93.92\% with Vi-SATNet re-generation, compared to 97.07\% to 9.83\% without re-generation. Furthermore, we introduce a visualization method to illustrate the rules learned by Vi-SATNets, thereby enhancing the interpretability of the pre-trained CNN model.
Supplementary Material: pdf
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.
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: 11151
Loading