Showing 1–4 of 4 results for author: Darwin, C

Do Counterfactual Examples Complicate Adversarial Training?

Authors: Eric Yeats, Cameron Darwin, Eduardo Ortega, Frank Liu, Hai Li

Abstract: We leverage diffusion models to study the robustness-performance tradeoff of robust classifiers. Our approach introduces a simple, pretrained diffusion method to generate low-norm counterfactual examples (CEs): semantically altered data which results in different true class membership. We report that the confidence and accuracy of robust models on their clean training data are associated with the… ▽ More We leverage diffusion models to study the robustness-performance tradeoff of robust classifiers. Our approach introduces a simple, pretrained diffusion method to generate low-norm counterfactual examples (CEs): semantically altered data which results in different true class membership. We report that the confidence and accuracy of robust models on their clean training data are associated with the proximity of the data to their CEs. Moreover, robust models perform very poorly when evaluated on the CEs directly, as they become increasingly invariant to the low-norm, semantic changes brought by CEs. The results indicate a significant overlap between non-robust and semantic features, countering the common assumption that non-robust features are not interpretable. △ Less

Submitted 17 April, 2024; v1 submitted 16 April, 2024; originally announced April 2024.

Comments: Accepted as a short paper to the GCV Workshop at CVPR'24

Adversarial Estimation of Topological Dimension with Harmonic Score Maps

Authors: Eric Yeats, Cameron Darwin, Frank Liu, Hai Li

Abstract: Quantification of the number of variables needed to locally explain complex data is often the first step to better understanding it. Existing techniques from intrinsic dimension estimation leverage statistical models to glean this information from samples within a neighborhood. However, existing methods often rely on well-picked hyperparameters and ample data as manifold dimension and curvature in… ▽ More Quantification of the number of variables needed to locally explain complex data is often the first step to better understanding it. Existing techniques from intrinsic dimension estimation leverage statistical models to glean this information from samples within a neighborhood. However, existing methods often rely on well-picked hyperparameters and ample data as manifold dimension and curvature increases. Leveraging insight into the fixed point of the score matching objective as the score map is regularized by its Dirichlet energy, we show that it is possible to retrieve the topological dimension of the manifold learned by the score map. We then introduce a novel method to measure the learned manifold's topological dimension (i.e., local intrinsic dimension) using adversarial attacks, thereby generating useful interpretations of the learned manifold. △ Less

Submitted 11 December, 2023; originally announced December 2023.

Comments: Accepted to the NeurIPS'23 Workshop on Diffusion Models

A quadratically enriched count of lines on a degree 4 del Pezzo surface

Authors: Cameron Darwin

Abstract: Over an algebraically closed field k, there are 16 lines on a degree 4 del Pezzo surface, but for other fields the situation is more subtle. In order to improve enumerative results over perfect fields, Kass and Wickelgren introduce a method analogous to counting zeroes of sections of smooth vector bundles using the Poincare-Hopf theorem. However, the technique of Kass-Wickelgren requires the enume… ▽ More Over an algebraically closed field k, there are 16 lines on a degree 4 del Pezzo surface, but for other fields the situation is more subtle. In order to improve enumerative results over perfect fields, Kass and Wickelgren introduce a method analogous to counting zeroes of sections of smooth vector bundles using the Poincare-Hopf theorem. However, the technique of Kass-Wickelgren requires the enumerative problem to satisfy a certain type of orientability condition. The problem of counting lines on a degree 4 del Pezzo surface does not satisfy this orientability condition, so most of the work of this paper is devoted to circumventing this problem. We do this by restricting to an open set where the orientability condition is satisfied, and checking that the count obtained is well-defined, similarly to an approach developed by Larson and Vogt. △ Less

Submitted 9 May, 2022; originally announced May 2022.

MSC Class: 14N15; 14F42

Conics meeting eight lines over perfect fields

Authors: Cameron Darwin, Aygul Galimova, Miao Pam Gu, Stephen McKean

Abstract: Over the complex numbers, there are 92 plane conics meeting 8 general lines in projective 3-space. Using the Euler class and local degree from motivic homotopy theory, we give an enriched version of this result over any perfect field. This provides a weighted count of the number of plane conics meeting 8 general lines, where the weight of each conic is determined the geometry of its intersections… ▽ More Over the complex numbers, there are 92 plane conics meeting 8 general lines in projective 3-space. Using the Euler class and local degree from motivic homotopy theory, we give an enriched version of this result over any perfect field. This provides a weighted count of the number of plane conics meeting 8 general lines, where the weight of each conic is determined the geometry of its intersections with the 8 given lines. As a corollary, real conics meeting 8 general lines come in two families of equal size. △ Less

Submitted 25 April, 2023; v1 submitted 12 July, 2021; originally announced July 2021.

Comments: 20 pages. Revised version with various errors corrected. Final version, but comments still welcome!

MSC Class: 14N15; 14F52

Journal ref: J. Algebra 631, 24 -- 45 (2023)