Showing 1–7 of 7 results for author: Campos, S

Verifying message-passing neural networks via topology-based bounds tightening

Authors: Christopher Hojny, Shiqiang Zhang, Juan S. Campos, Ruth Misener

Abstract: Since graph neural networks (GNNs) are often vulnerable to attack, we need to know when we can trust them. We develop a computationally effective approach towards providing robust certificates for message-passing neural networks (MPNNs) using a Rectified Linear Unit (ReLU) activation function. Because our work builds on mixed-integer optimization, it encodes a wide variety of subproblems, for exam… ▽ More Since graph neural networks (GNNs) are often vulnerable to attack, we need to know when we can trust them. We develop a computationally effective approach towards providing robust certificates for message-passing neural networks (MPNNs) using a Rectified Linear Unit (ReLU) activation function. Because our work builds on mixed-integer optimization, it encodes a wide variety of subproblems, for example it admits (i) both adding and removing edges, (ii) both global and local budgets, and (iii) both topological perturbations and feature modifications. Our key technology, topology-based bounds tightening, uses graph structure to tighten bounds. We also experiment with aggressive bounds tightening to dynamically change the optimization constraints by tightening variable bounds. To demonstrate the effectiveness of these strategies, we implement an extension to the open-source branch-and-cut solver SCIP. We test on both node and graph classification problems and consider topological attacks that both add and remove edges. △ Less

Submitted 21 May, 2024; v1 submitted 21 February, 2024; originally announced February 2024.

Optimizing over trained GNNs via symmetry breaking

Authors: Shiqiang Zhang, Juan S. Campos, Christian Feldmann, David Walz, Frederik Sandfort, Miriam Mathea, Calvin Tsay, Ruth Misener

Abstract: Optimization over trained machine learning models has applications including: verification, minimizing neural acquisition functions, and integrating a trained surrogate into a larger decision-making problem. This paper formulates and solves optimization problems constrained by trained graph neural networks (GNNs). To circumvent the symmetry issue caused by graph isomorphism, we propose two types o… ▽ More Optimization over trained machine learning models has applications including: verification, minimizing neural acquisition functions, and integrating a trained surrogate into a larger decision-making problem. This paper formulates and solves optimization problems constrained by trained graph neural networks (GNNs). To circumvent the symmetry issue caused by graph isomorphism, we propose two types of symmetry-breaking constraints: one indexing a node 0 and one indexing the remaining nodes by lexicographically ordering their neighbor sets. To guarantee that adding these constraints will not remove all symmetric solutions, we construct a graph indexing algorithm and prove that the resulting graph indexing satisfies the proposed symmetry-breaking constraints. For the classical GNN architectures considered in this paper, optimizing over a GNN with a fixed graph is equivalent to optimizing over a dense neural network. Thus, we study the case where the input graph is not fixed, implying that each edge is a decision variable, and develop two mixed-integer optimization formulations. To test our symmetry-breaking strategies and optimization formulations, we consider an application in molecular design. △ Less

Submitted 12 October, 2023; v1 submitted 16 May, 2023; originally announced May 2023.

Comments: 10 main pages, 27 with appendix, 10 figures, 7 tables

The H2-optimal Control Problem of CSVIU Systems: Discounted, Counter-discounted and Long-run Solutions -- Part II: Optimal Control

Authors: João B. R. do Val, Daniel S. Campos

Abstract: The paper deals with stochastic control problems associated with $H_2$ performance indices such as energy or power norms or energy measurements when norms are not defined. They apply to a class of systems for which a stochastic process conveys the underlying uncertainties, known as CSVIU (Control and State Variation Increase Uncertainty). These indices allow various emphases from focusing on the t… ▽ More The paper deals with stochastic control problems associated with $H_2$ performance indices such as energy or power norms or energy measurements when norms are not defined. They apply to a class of systems for which a stochastic process conveys the underlying uncertainties, known as CSVIU (Control and State Variation Increase Uncertainty). These indices allow various emphases from focusing on the transient behavior with the discounted norm to stricter conditions on stability, steady-state mean-square error and convergence rate, using the optimal overtaking criterion -- the long-run average power control stands as a midpoint in this respect. A critical advance regards the explicit form of the optimal control law, expressed in two equivalent forms. One takes a perturbed affine Riccati-like form of feedback solution; the other comes from a generalized normal equation that arises from the nondifferentiable local optimal problem. They are equivalent, but the latter allows for a search method to attain the optimal law. A detectability notion and a Riccati solution grant stochastic stability from the behavior of the norms. The energy overtaking criterion requires a further constraint on a matrix spectral radius. With these findings, the paper revisits the emerging of the inaction solution, a prominent feature of CSVIU models to deal with the uncertainty inherent to poorly known models. Besides, it provides the optimal solution and the tools to pursue it. △ Less

Submitted 25 June, 2021; originally announced June 2021.

The H2-optimal Control Problem of CSVIU Systems: Discounted, Counter-discounted and Long-Run Solutions -- Part I: The Norm

Authors: João B. R. do Val, Daniel S. Campos

Abstract: The paper deals with the H2-norm and associated energy or power measurements for a class of processes known as CSVIU (Control and State Variation Increase Uncertainty). These are system models for which a stochastic process conveys the underlying uncertainties, and are able to give rise to cautious controls. The paper delves into the non-controlled version and fundamental system and norms notions… ▽ More The paper deals with the H2-norm and associated energy or power measurements for a class of processes known as CSVIU (Control and State Variation Increase Uncertainty). These are system models for which a stochastic process conveys the underlying uncertainties, and are able to give rise to cautious controls. The paper delves into the non-controlled version and fundamental system and norms notions associated with stochastic stability and mean-square convergence. One pillar of the study is the connection between the finiteness of one of these norms or a limited energy measurement growth with the corresponding stochastic stability notions. A detectability concept ties these notions, and the analysis of linear-positive operators plays a fundamental role. The introduction of various H2-norms and energy measurement performance criteria allows one to span the focus from transient to long-run behavior. As the discount parameter turns into a counter-discount, the criteria enforce stricter requirements on the second-moment steady state errors and on the exponential convergence rate. A tidy connection among this H2-performance measures cast employs a unifying vanishing discount reasoning. △ Less

Submitted 25 June, 2021; originally announced June 2021.

On the number of sets with a given doubling constant

Authors: Marcelo Soares Campos

Abstract: We study the number of $s$-element subsets $J$ of a given abelian group $G$, such that $|J+J|\leq K|J|$. Proving a conjecture of Alon, Balogh, Morris and Samotij, and improving a result of Green and Morris, who proved the conjecture for $K$ fixed, we provide an upper bound on the number of such sets which is tight up to a factor of $2^{o(s)},$ when $G=\mathbb{Z}$ and $K=o(s/(\log n)^3)$. We also p… ▽ More We study the number of $s$-element subsets $J$ of a given abelian group $G$, such that $|J+J|\leq K|J|$. Proving a conjecture of Alon, Balogh, Morris and Samotij, and improving a result of Green and Morris, who proved the conjecture for $K$ fixed, we provide an upper bound on the number of such sets which is tight up to a factor of $2^{o(s)},$ when $G=\mathbb{Z}$ and $K=o(s/(\log n)^3)$. We also provide a generalization of this result to arbitrary abelian groups which is tight up to a factor of $2^{o(s)}$ in many cases. The main tool used in the proof is the asymmetric container lemma, introduced recently by Morris, Samotij and Saxton. △ Less

Submitted 3 May, 2019; v1 submitted 13 November, 2018; originally announced November 2018.

Comments: 11 pages + 10 page appendix. Minor mistakes and typos from the first version were corrected. arXiv admin note: substantial text overlap with arXiv:1806.03706 by other authors

Exceptional sets for nonuniformly hyperbolic diffeomorphisms

Authors: Sara Campos, Katrin Gelfert

Abstract: For a surface diffeomorphism, a compact invariant locally maximal set $W$ and some subset $A\subset W$ we study the $A$-exceptional set, that is, the set of points whose orbits do not accumulate at $A$. We show that if the Hausdorff dimension of $A$ is smaller than the Hausdorff dimension $d$ of some ergodic hyperbolic measure, then the topological entropy of the exceptional set is at least the en… ▽ More For a surface diffeomorphism, a compact invariant locally maximal set $W$ and some subset $A\subset W$ we study the $A$-exceptional set, that is, the set of points whose orbits do not accumulate at $A$. We show that if the Hausdorff dimension of $A$ is smaller than the Hausdorff dimension $d$ of some ergodic hyperbolic measure, then the topological entropy of the exceptional set is at least the entropy of this measure and its Hausdorff dimension is at least $d$. Particular consequences occur when there is some a priori defined hyperbolic structure on $W$ and, for example, if there exists an SRB measure. △ Less

Submitted 29 December, 2017; originally announced January 2018.

Exceptional sets for nonuniformly expanding maps

Authors: Sara Campos, Katrin Gelfert

Abstract: Given a rational map of the Riemann sphere and a subset $A$ of its Julia set, we study the $A$-exceptional set, that is, the set of points whose orbit does not accumulate at $A$. We prove that if the topological entropy of $A$ is less than the topological entropy of the full system then the $A$-exceptional set has full topological entropy. Furthermore, if the Hausdorff dimension of $A$ is smaller… ▽ More Given a rational map of the Riemann sphere and a subset $A$ of its Julia set, we study the $A$-exceptional set, that is, the set of points whose orbit does not accumulate at $A$. We prove that if the topological entropy of $A$ is less than the topological entropy of the full system then the $A$-exceptional set has full topological entropy. Furthermore, if the Hausdorff dimension of $A$ is smaller than the dynamical dimension of the system then the Hausdorff dimension of the $A$-exceptional set is larger than or equal to the dynamical dimension, with equality in the particular case when the dynamical dimension and the Hausdorff dimension coincide. We discuss also the case of a general conformal $C^{1+α}$ dynamical system and, in particular, certain multimodal interval maps on their Julia set. △ Less

Submitted 15 February, 2016; v1 submitted 1 June, 2015; originally announced June 2015.