Tropical Geometry Features for Novelty Detection and interpretability

Anupam Chaudhuri, Anj Simmons

Keywords: OOD, Tropical geometry, Deep neural network, polytopes

Abstract:

Existing methods for critical tasks such as out-of-distribution (OOD) detection,
uncertainty quantification, and adversarial robustness often focus on measuring
the output of the last or intermediate layers of a neural network such as logits and
energy score. However, these methods typically overlook the geometric properties
of the learned representations in the latent space, failing to capture important
signals that relate to model reliability, fairness, and adversarial vulnerability.

Innovations: We introduce an innovative method, termed Tropical Geometry Features (TGF), for detecting out-of-distribution data and enhancing overall model evaluation. This approach leverages the geometric properties of polytopes derived
from a trained neural network’s learned representations. By integrating these
geometric features with the data used during training, TGF establishes a unique
signature of in-distribution data points. Our framework extends beyond OOD
detection, providing insights into model uncertainty, adversarial robustness, interpretability, and fairness. Through TGF, we enhance interpretability technique to detect OOD, uncertainty, adverserial robustness in dynamic and unpredictable
environments.

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: 5570