Simpler PAC-Bayesian bounds for hostile data

PAC-Bayesian learning bounds are of the utmost interest to the learning community. Their role is to connect the generalization ability of an aggregation distribution \(\rho \) to its empirical risk and to its Kullback-Leibler divergence with respect to some prior distribution \(\pi \). Unfortunately, most of the available bounds typically rely on heavy assumptions such as boundedness and independence of the observations. This paper aims at relaxing these constraints and provides PAC-Bayesian learning bounds that hold for dependent, heavy-tailed observations (hereafter referred to as hostile data). In these bounds the Kullack-Leibler divergence is replaced with a general version of Csiszár’s f-divergence. We prove a general PAC-Bayesian bound, and show how to use it in various hostile settings.     

Population computation of vectorial transformations

Many neurons of the central nervous system are broadly tuned to some sensory or motor variables. This property allows one to assign to each neuron a preferred attribute (PA). The width of tuning curves and the distribution of PAs in a population of neurons tuned to a given variable define the collective behavior of the population. In this article, we study the relationship of the nature of the tuning curves, the distribution of PAs, and computational properties of linear neuronal populations. We show that noise-resistant distributed linear algebraic processing and learning can be implemented by a population of cosine tuned neurons assuming a nonuniform but regular distribution of PAs. We extend these results analytically to the noncosine tuning and uniform distribution case and show with a numerical simulation that the results remain valid for a nonuniform regular distribution of PAs for broad noncosine tuning curves. These observations provide a theoretical basis for modeling general nonlinear sensorimotor transformations as sets of local linearized representations.     

An HLLC Scheme to Solve The <Emphasis Type="Italic">M</Emphasis><Subscript>1</Subscript> Model of Radiative Transfer in Two Space Dimensions

The M ₁ radiative transfer model is considered in the present work in order to simulate the radiative fields and their interactions with the matter. The model is governed by an hyperbolic system of conservation laws supplemented by relaxation source terms. Several difficulties arise when approximating the solutions of the model; namely the positiveness of the energy, the flux limitation and and the limit diffusion behavior have to be satisfied. An HLLC scheme is exhibited and it is shown to satisfy all the required properties. A particular attention is payed concerning the approximate extreme waves. These approximations are crucial to obtain an accurate scheme. The extension to the full 2D problem is proposed. It satisfies, once again, all the expected properties. Numerical experiments are proposed. They show that the considered scheme is actually less diffusive than the currently used numerical methods.      

Kinetic study of selenium used as optical data storage and holographic recording materials

A kinetic study has been performed on selenium for isothermal and continuous heating processes. The activation energy (E _a) and Avrami exponent (n) have been evaluated in the range 414 to 449 K for the isothermal process and at a heating rate of 5 to 40 K min^–1 for continuous heating. A new iterative calculation to give exact values ofE _a andn for a continuous heating process is also presented.     

An algorithm for spectral factorization using random search techniques

An approximation method for the energy spectrum of a stationary stochastic dynamical system is presented, which allows approximate functional rational factorization.This paper is in three parts. The first deals with a theoretical problem of approximation in Hardy Spaces, whose main result is the following:Let S⁽ⁱⁿ⁾, S be positive functions belonging to L¹(R_gw), such that log S⁽ⁿ⁾ and log S belong to L¹(R_ω, dω/1 + ω²).Let h⁽ⁿ⁾, h be the outer functions of the Hardy Space H²(Π₊) such that S^(g) = |h⁽ⁿ⁾|² and S = |h|² on iR.If S⁽ⁿ⁾ n^→∝ S in L¹(R_ω), and log S⁽ⁿ⁾ n^→∝ log S in L¹(R_ω, dω/1 + ω²), then: h⁽ⁿ⁾ n^→∝ h in H²(Π₊).The second part describes an effective algorithm, using random search methods, and gives an almost sure convergence result for it.The third part treats numerically two examples, permitting comparison of this algorithm with others (whenever there are…): the first example is a problem of approximation for a nonrational process (turbulence) that was considered in Ref. 22: the second example is a problem of model reduction (automatic) considered in Ref. 4.     

Sense induction in folksonomies: a review

Folksonomies, often known as tagging systems, such as the ones used on the popular Delicious or Flickr websites, use a very simple Knowledge Organisation System. Users have thus been quick to adopt this system and create extensive annotations on the Web. However, because of the simplicity of the folksonomy model, the semantics of the tags used is not explicit and can only be inferred from their context of use. This is a barrier for the automatic use of such Knowledge Organisation Systems by computers and new techniques have been developed to extract the semantic of the tags. In this article we discuss the drawbacks of some of these approaches and propose a generalization of the different approaches to detect new senses of terms in a folksonomy. Another weak point of the current state of the art in the field is the lack of formal evaluation methodology; we thus propose a novel evaluation framework. We introduce a dataset and evaluation methodology that enable the comparison of results between different approaches to sense induction in folksonomies. Finally we discuss the performances of different approaches to the task of homonymous/polysemous tag detection and synonymous identification.     

The complexity of intersecting finite automata having few final states

The problem of determining whether several finite automata accept a word in common is closely related to the well-studied membership problem in transformation monoids. We raise the issue of limiting the number of final states in the automata intersection problem. For automata with two final states, we show the problem to be \({\oplus}\)L-complete or NP-complete according to whether a nontrivial monoid other than a direct product of cyclic groups of order 2 is allowed in the automata. We further consider idempotent commutative automata and (Abelian, mainly) group automata with one, two, or three final states over a singleton or larger alphabet, elucidating (under the usual hypotheses on complexity classes) the complexity of the intersection nonemptiness and related problems in each case.     

GEAMAS: A Generic Architecture for Agent-Oriented Simulations of Complex Processes

This paper's object is to present the results of the GEAMAS project which aims at modeling and simulating natural complex systems. GEAMAS is a generic architecture of agents used to study the behavior emergence in such systems. It is a multiagent program meant to develop simulation applications. Modeling complex systems requires to reduce, to organize the system complexity and to describe suitable components. Complexity of the system can then be tackled with an agent-oriented approach, where interactions lead to a global behavior. This approach helps in understanding how non-determinist behavior can emerge from interactions between agents, which is near of self-organized criticality used to explain natural phenomena. In the Applied Artificial Intelligence context, this paper presents an agent software architecture using a model of agent. This architecture is composed of three abstract levels over which the complexity is distributed and reduced. The architecture is implemented in ReActalk, an open agent-oriented development tool, which was developed on top of Smalltalk-80. To illustrate our purpose and to validate the architecture, a simulation program to help in predicting volcanic eruptions was investigated. This program was run over a period of one year and has given many satisfying results unattainable up to there with more classical approaches.     

Particle-based Liquid Control using Animation Templates

It is notoriously difficult for artists to control liquids while generating plausible animations. We introduce a new liquid control tool that allows users to load, transform, and apply precomputed liquid simulation templates in a scene in order to control a particle-based simulation. Each template instance generates control forces that drive the global simulated liquid to locally reproduce the templated liquid behavior. Our system is augmented with a variable proportion of temporary particles to help efficiently reproduce the templated liquid density, with fewer requirements on the surrounding environment. The resulting control strategy adds only a small computational overhead, leading to quick visual feedback for resolutions allowing interactive simulation. We demonstrate the robustness and ease of use of our method on various examples in 2D and 3D.