Search

Most data poisoning methods target naive deep networks. Yet, it is well known that those networks exhibit strong sensitivity to perturbations. Inversely, in this paper, I show that several data poisoning attacks (e.g. poison frog) are ineffective as soon as there are applied on features made robust to adversarial attacks, on both CIFAR and MNIST datasets. This result stresses that some state of the art data poisoning results may have been corrupted by adversarial sensibility and should be further checked on robust networks instead of naive ones. Code is available at github.com/achanhon/AdversarialModel/V4.

International audience; Today, adversarial attack is almost the only kind of hard samples considered by the community to tickle deep networks.However, such additive perturbations are not realistic for many applications (e.g. remote sensing).This paper introduce a new kind of hard samples called corner cases where the entire data is generated (not just a perturbation). This allows to constraint more easily those data to be both hard but also realistic.

Gaussian elimination is known to be exponential when done naively. Indeed, theoretically, it is required to take care of the intermediary numbers encountered during an algorithm, in particular of their binary sizes. However, this point is weakly tackled for linear programming in state of the art.Thus, this paper introduces a new polynomial times algorithm for linear programming focusing on this point: this algorithm offers an explicit strategy to deal with all number representation issues. One key feature which makes this Newton based algorithm more compliant with binary considerations is that the optimization is performed in the so-called first phase of Newton descent and not in the so-called second phase like in the state of the art.

This paper proves the existence of a vertex based algorithm, and, an interior point algorithm which can help each other while solving linear programming.This result is not trivial as, in the state of the art, those two kind of algorithms have incompatible ways to reach a solution.Importantly, these two new algorithms have the property expected for interior point/vertex based algorithms.The interior point algorithm is polynomial (with a complexity only slightly higher than the state of the art), and, the number of steps of the vertex based algorithm is related to a number of vertices.

This document presents two algorithms for linear programming which seems to be able to work together while one is combinatorial (number of iterations is lower than number of subset of the constraint set) and the other is interior (number of iteration is polynomial but depend on the binary size of the input).

This paper offers a new polynomial algorithm for linear programming liked to Perceptron.Despite that the arithmetic complexity of this algorithm is slightly higher than path-following one (current state of the art), the offered algorithm is suitable for exact computation.This key feature is made possible thank to an explicit (and simple) rounding process and an explicit integer transformation on matrices manipulated during computation.

This paper offers a new polynomial algorithm for linear feasibility.Despite, arithmetic complexity of this algorithm is slightly higher than best state of the art path-following one, the offered algorithm is very suitable for exact computation as simple transformation allows to handle integer matrix.All bounds required for establishing complexity have simple explicit values (except for the number of steps classically linked with a sub determinant of the input matrix).In particular, rounding process is explicit.

It is well known that minimization of a self concordant function has two phases.Literature mainly focuses on the so called second phase (which starts when the Newton increment is lower than a critical value) where optimization exhibits a quadratic behavior.Yet, paper hal-02491694 has found an interesting algorithm for linear programming where (almost) all the Newton descent is in first phase setting.This invites to take a new look on the first phase of this algorithm.

Naive Gaussian elimination is known to be exponential when done with arbitrary large integer representation. This highlights that binary properties of algorithms are (at least theoretically) important.Yet, state of the art of linear programming only weakly tackles this issue.Indeed, this paper stresses some difficulties which can arise when solving linear programming with classical methods under arbitrary large integer representation setting.Then, it introduces a new polynomial times algorithm for linear programming with better binary properties.

Adversarial examples are a burring issue for deep learning.Yet, the basic definition of adversarial example is not that trivial.Indeed, if one consider adversarial examples as points close to decision boundary, then, one can prove that adversarial examples are not only inevitable but almost everywhere.Yet, seeing adversarial examples as points not expected to be close to decision boundary may lead to very different conclusions.

This paper presents a new algorithm for linear feasibility which requires at most $\widetilde{O}(L)$ damped Newton steps to solve a linear feasibility instance of total binary size $L$. This is slightly less than usual log barrier way ($\widetilde{O}(\sqrt{N}L)$ Newton steps), Chubanov's way ($\widetilde{O}(N^4L)$ total time complexity) or rescaling way ($\widetilde{O}(\sqrt{M}NL)$ Perceptron steps).In addition, this algorithm, which encodes the original Perceptron idea into a self concordant function, does not directly rely on central-path framework. This feature may be of interest in view of the recent negative results for this framework.

This document presents an algorithm for linear programming whose complexity seems to be O(M^3L) with a main loop which is just a simple coordinate descent. This results, which should be validated, is interesting for linear program.

This paper introduces an algorithms for linear feasibility which is deterministic, exact, and, with a linear complexity (in step - each step corresponding to solve a linear system).In addition, this algorithm has several features which should be explored. First, both the numbers of steps with high or low Newton increment are negligible. Then, this algorithm does not seem to be related to central path when extended to linear programming.

This paper is a summary of paper hal-02491694 focusing on the main theorem.

International audience; Nous présentons une approche interactive pour la segmentation multi-classe d’images aériennes. Celle-ci repose sur un réseau de neurones profond qui exploite à la fois des images RGB et des annotations. À partir d’une sortie initiale qui dépend uniquement de l’image, notre réseau affine ensuite de manière interactive ce résultat en utilisant une concaténation de l’image et des annotations de l’utilisateur. Il est important de noter qu’il n’y a pas de réentraînement durant les interactions, ce qui permet un processus rapide et flexible. Dans le contexte d’images extrêmement résolues telles que celles acquises par drone, nous montrons qu’un humain-dans-la-boucle est très gratifiant, quel que soit l’environnement expérimental et l’architecture du réseau utilisée.

International audience; Attention can be oriented in space covertly without the need of eye movements. We used multivariate pattern classification analyses (MVPA) to investigate whether the time course of the deployment of covert spatial attention leading up to the observer’s perceptual decision can be decoded from both EEG alpha power and raw activity traces. Decoding attention from these signals can help determine whether raw EEG signals and alpha power reflect the same or distinct features of attentional selection. Using a classical cueing task, we showed that the orientation of covert spatial attention can be decoded by both signals. However, raw activity and alpha power may reflect different features of spatial attention, with alpha power more associated with the orientation of covert attention in space and raw activity with the influence of attention on perceptual processes.

State-of-the-art log barrier algorithms for linear programming rely on central-path framework.Recently, some hard instances for any algorithms relying on this framework has been found, making relevant to look for not-central-path alternatives.This paper introduces a simple log barrier algorithm with similar times complexity than previous ones, but, with a different underlying mechanism.Currently, this algorithm could be seen as a log barrier extension of the Perceptron, and, has some interesting features for machine learning context.

International audience; Artificial Intelligence and Decision Making Challenges in the Certification of Computer Vision-Based Systems for Civil Aeronautics C omputer vision techniques have made considerable progress in recent years. This advance now makes possible the practical use of computer vision in civil drones or aircraft, replacing human pilots. The question that naturally arises is then to provide a way to certify those types of systems at a given level of safety. The aim of the article is, firstly, to understand the gap between today's computer vision systems and the current certification standards; secondly, to identify the key activities that must be fulfilled in order to make computer-vision systems certifiable and, thirdly, to explore some recent works related to these key activities.

In this paper, we introduce a framework composed of two algorithms for linear programming.The first algorithm is interior point, log barrier, and, polynomial times.The second one is probably exponential times, but, not depending on the binary size of numbers, and, combinatorial (i.e. current point can be matched to a set of constraints/variables).And, the interest is that each algorithms could help the other as they work on similar space.

Minimizing $F(v)=v^TAA^Tv - \overset{N}{\underset{n=1}{\sum}}\log(v_n)$ allows to produce a solution of $Ax>\mathbf{0}$.As $F(v)$ is self concordant, this minimization can be done using Newton descent leading to a polynomial times algorithm for linear programming.But, contrary to classical log barrier, Newton only offers a constant decrease on $F$. Thus, intertwining Newton step with other strategy could be interesting.

Minimizing $F(v)=v^TAA^Tv - \overset{N}{\underset{n=1}{\sum}}\log(v_n)$ allows to produce a solution of $Ax>\mathbf{0}$.As $F(v)$ is self concordant, this minimization can be done using Newton descent, and, leads to a polynomial times algorithm without any numerical issue.Interestingly, this descent can be hybridized with other minimization strategy like approximating $F$ as a sum of independent functions which seems interesting during the first phase of optimization.

Minimizing $F(v)=v^TAA^Tv - \overset{N}{\underset{n=1}{\sum}}\log(v_n)$ allows to produce a solution of $Ax>\mathbf{0}$.As $F(v)$ is self concordant, this minimization can be done using Newton descent, and, leads to a polynomial times algorithm.Interestingly, this descent can be hybridized with other minimization strategy. Despite naive ways (gradient step, Perceptron, coordinate based) are probably useless, approximating $F$ as a sum of independent function seems interesting during the first phase of optimization.

This paper focuses on different ways to solve the system of non linear equations $AA^Tv=\frac{1}{v}$ (allowing to find $x$ such that $Ax>\mathbf{0}$ itself equivalent to linear programming).

Homogeneous linear feasibility is polynomially equivalent to linear programming. But, contrary to original linear program, homogeneous linear feasibility can be tackled in two ways: unstructured or structured. This paper presents the pros/cons of these two ways for both simplex and log barrier algorithms: structured version being less efficient alone but more suitable for hybridization.