Distributed conditional computation

L'objectif de cette thèse est de présenter différentes applications du programme de recherche de calcul conditionnel distribué. On espère que ces applications, ainsi que la théorie présentée ici, mènera à une solution générale du problème d'intelligence artificielle, en particulier en ce qui a trait à la nécessité d'efficience. La vision du calcul conditionnel distribué consiste à accélérer l'évaluation et l'entraînement de modèles profonds, ce qui est très différent de l'objectif usuel d'améliorer sa capacité de généralisation et d'optimisation. Le travail présenté ici a des liens étroits avec les modèles de type mélange d'experts. Dans le chapitre 2, nous présentons un nouvel algorithme d'apprentissage profond qui utilise une forme simple d'apprentissage par renforcement sur un modèle d'arbre de décisions à base de réseau de neurones. Nous démontrons la nécessité d'une contrainte d'équilibre pour maintenir la distribution d'exemples aux experts uniforme et empêcher les monopoles. Pour rendre le calcul efficient, l'entrainement et l'évaluation sont contraints à être éparse en utilisant un routeur échantillonnant des experts d'une distribution multinomiale étant donné un exemple. Dans le chapitre 3, nous présentons un nouveau modèle profond constitué d'une représentation éparse divisée en segments d'experts. Un modèle de langue à base de réseau de neurones est construit à partir des transformations éparses entre ces segments. L'opération éparse par bloc est implémentée pour utilisation sur des cartes graphiques. Sa vitesse est comparée à deux opérations denses du même calibre pour démontrer le gain réel de calcul qui peut être obtenu. Un modèle profond utilisant des opérations éparses contrôlées par un routeur distinct des experts est entraîné sur un ensemble de données d'un milliard de mots. Un nouvel algorithme de partitionnement de données est appliqué sur un ensemble de mots pour hiérarchiser la couche de sortie d'un modèle de langage, la rendant ainsi beaucoup plus efficiente. Le travail présenté dans cette thèse est au centre de la vision de calcul conditionnel distribué émis par Yoshua Bengio. Elle tente d'appliquer la recherche dans le domaine des mélanges d'experts aux modèles profonds pour améliorer leur vitesse ainsi que leur capacité d'optimisation. Nous croyons que la théorie et les expériences de cette thèse sont une étape importante sur la voie du calcul conditionnel distribué car elle cadre bien le problème, surtout en ce qui concerne la compétitivité des systèmes d'experts.

Information Discovery in Distributed Systems-A Novel Concept Based on Modelling of Information Elements

Sharing of information with those in need of it has always been an idealistic goal of networked environments. With the proliferation of computer networks, information is so widely distributed among systems, that it is imperative to have well-organized schemes for retrieval and also discovery. This thesis attempts to investigate the problems associated with such schemes and suggests a software architecture, which is aimed towards achieving a meaningful discovery. Usage of information elements as a modelling base for efficient information discovery in distributed systems is demonstrated with the aid of a novel conceptual entity called infotron.The investigations are focused on distributed systems and their associated problems. The study was directed towards identifying suitable software architecture and incorporating the same in an environment where information growth is phenomenal and a proper mechanism for carrying out information discovery becomes feasible. An empirical study undertaken with the aid of an election database of constituencies distributed geographically, provided the insights required. This is manifested in the Election Counting and Reporting Software (ECRS) System. ECRS system is a software system, which is essentially distributed in nature designed to prepare reports to district administrators about the election counting process and to generate other miscellaneous statutory reports.Most of the distributed systems of the nature of ECRS normally will possess a "fragile architecture" which would make them amenable to collapse, with the occurrence of minor faults. This is resolved with the help of the penta-tier architecture proposed, that contained five different technologies at different tiers of the architecture.The results of experiment conducted and its analysis show that such an architecture would help to maintain different components of the software intact in an impermeable manner from any internal or external faults. The architecture thus evolved needed a mechanism to support information processing and discovery. This necessitated the introduction of the noveI concept of infotrons. Further, when a computing machine has to perform any meaningful extraction of information, it is guided by what is termed an infotron dictionary.The other empirical study was to find out which of the two prominent markup languages namely HTML and XML, is best suited for the incorporation of infotrons. A comparative study of 200 documents in HTML and XML was undertaken. The result was in favor ofXML.The concept of infotron and that of infotron dictionary, which were developed, was applied to implement an Information Discovery System (IDS). IDS is essentially, a system, that starts with the infotron(s) supplied as clue(s), and results in brewing the information required to satisfy the need of the information discoverer by utilizing the documents available at its disposal (as information space). The various components of the system and their interaction follows the penta-tier architectural model and therefore can be considered fault-tolerant. IDS is generic in nature and therefore the characteristics and the specifications were drawn up accordingly. Many subsystems interacted with multiple infotron dictionaries that were maintained in the system.In order to demonstrate the working of the IDS and to discover the information without modification of a typical Library Information System (LIS), an Information Discovery in Library Information System (lDLIS) application was developed. IDLIS is essentially a wrapper for the LIS, which maintains all the databases of the library. The purpose was to demonstrate that the functionality of a legacy system could be enhanced with the augmentation of IDS leading to information discovery service. IDLIS demonstrates IDS in action. IDLIS proves that any legacy system could be augmented with IDS effectively to provide the additional functionality of information discovery service.Possible applications of IDS and scope for further research in the field are covered.

Systematic trends in self-consistent calculations of linear quantum wires

Systematic trends in the properties of a linear split-gate heterojunction are studied by solving iteratively the Poisson and Schrödinger equations for different gate potentials and temperatures. A two-dimensional approximation is presented that is much simpler in the numerical implementation and that accurately reproduces all significant trends. In deriving this approximation, we provide a rigorous and quantitative basis for the formulation of models that assumes a two-dimensional character for the electron gas at the junction.

Density-functional calculations of magnetoplasmons in quantum rings

We have studied the structure and dipole charge-density response of nanorings as a function of the magnetic field using local-spin-density-functional theory. Two small rings consisting of 12 and 22 electrons confined by a positively charged background are used to represent the cases of narrow and wide rings. The results are qualitatively compared with experimental data existing on microrings and on antidots. A smaller ring containing five electrons is also analyzed to allow for a closer comparison with a recent experiment on a two-electron quantum ring.

Helium in polygonal nanopores at zero temperature: Density functional theory calculations

We investigate adsorption of helium in nanoscopic polygonal pores at zero temperature using a finite-range density functional theory. The adsorption potential is computed by means of a technique denoted as the elementary source method. We analyze a rhombic pore with Cs walls, where we show the existence of multiple interfacial configurations at some linear densities, which correspond to metastable states. Shape transitions and hysterectic loops appear in patterns which are richer and more complex than in a cylindrical tube with the same transverse area.

Information discovery in distributed systems - A novel concept based on modelling of information elements

This thesis attempts to investigate the problems associated with such schemes and suggests a software architecture, which is aimed towards achieving a meaningful discovery. Usage of information elements as a modelling base for efficient information discovery in distributed systems is demonstrated with the aid of a novel conceptual entity called infotron. The investigations are focused on distributed systems and their associated problems. The study was directed towards identifying suitable software architecture and incorporating the same in an environment where information growth is phenomenal and a proper mechanism for carrying out information discovery becomes feasible. An empirical study undertaken with the aid of an election database of constituencies distributed geographically, provided the insights required. This is manifested in the Election Counting and Reporting Software (ECRS) System. ECRS system is a software system, which is essentially distributed in nature designed to prepare reports to district administrators about the election counting process and to generate other miscellaneous statutory reports.

Performance of correlation functionals in ab initio chemisorption cluster-model calculations: Alkali metals on Si(111)

The performance of different correlation functionals has been tested for alkali metals, Li to Cs, interacting with cluster models simulating different active sites of the Si(111) surface. In all cases, the ab initio Hartree-Fock density has been obtained and used as a starting point. The electronic correlation energy is then introduced as an a posteriori correction to the Hartree-Fock energy using different correlation functionals. By making use of the ionic nature of the interaction and of different dissociation limits we have been able to prove that all functionals tested introduce the right correlation energy, although to a different extent. Hence, correlation functionals appear as an effective and easy way to introduce electronic correlation in the ab initio Hartree-Fock description of the chemisorption bond in complex systems where conventional configuration interaction techniques cannot be used. However, the calculated energies may differ by some tens of eV. Therefore, these methods can be employed to get a qualitative idea of how important correlation effects are, but they have some limitations if accurate binding energies are to be obtained.

Magnetic structure of Li2CuO2: From ab initio calculations to macroscopic simulations

The magnetic structure of the edge-sharing cuprate compound Li2CuO2 has been investigated with highly correlated ab initio electronic structure calculations. The first- and second-neighbor in-chain magnetic interactions are calculated to be 142 and -22 K, respectively. The ratio between the two parameters is smaller than suggested previously in the literature. The interchain interactions are antiferromagnetic in nature and of the order of a few K only. Monte Carlo simulations using the ab initio parameters to define the spin model Hamiltonian result in a Nel temperature in good agreement with experiment. Spin population analysis situates the magnetic moment on the copper and oxygen ions between the completely localized picture derived from experiment and the more delocalized picture based on local-density calculations.

Effective t-J model Hamiltonian parameters of monolayered cuprate superconductors from ab initio electronic structure calculations

The magnetic coupling constant of selected cuprate superconductor parent compounds has been determined by means of embedded cluster model and periodic calculations carried out at the same level of theory. The agreement between both approaches validates the cluster model. This model is subsequently employed in state-of-the-art configuration interaction calculations aimed to obtain accurate values of the magnetic coupling constant and hopping integral for a series of superconducting cuprates. Likewise, a systematic study of the performance of different ab initio explicitly correlated wave function methods and of several density functional approaches is presented. The accurate determination of the parameters of the t-J Hamiltonian has several consequences. First, it suggests that the appearance of high-Tc superconductivity in existing monolayered cuprates occurs with J/t in the 0.20¿0.35 regime. Second, J/t=0.20 is predicted to be the threshold for the existence of superconductivity and, third, a simple and accurate relationship between the critical temperatures at optimum doping and these parameters is found. However, this quantitative electronic structure versus Tc relationship is only found when both J and t are obtained at the most accurate level of theory.

Coherent tunnelling in Cu2+- and Ag2+-doped MgO and CaO: Cu2+ explored through ab initio calculations

The observation of coherent tunnelling in Cu2+ - and Ag2+ -doped MgO and CaO:Cu2+ was a crucial discovery in the realm of the Jahn-Teller (JT) effect. The main reasons favoring this dynamic behavior are now clarified through ab initio calculations on Cu2+ - and Ag2+ -doped cubic oxides. Small JT distortions and an unexpected low anharmonicity of the eg JT mode are behind energy barriers smaller than 25 cm-1 derived through CASPT2 calculations for Cu2+ - and Ag2+ -doped MgO and CaO:Cu2+ . The low anharmonicity is shown to come from a strong vibrational coupling of MO610- units (M=Cu,Ag) to the host lattice. The average distance between the d9 impurity and ligands is found to vary significantly on passing from MgO to SrO following to a good extent the lattice parameter.

Channel Adaptive MAC Protocol with Traffic-Aware Distributed Power Management in Wireless Sensor Networks-Some Performance Issues

In Wireless Sensor Networks (WSN), neglecting the effects of varying channel quality can lead to an unnecessary wastage of precious battery resources and in turn can result in the rapid depletion of sensor energy and the partitioning of the network. Fairness is a critical issue when accessing a shared wireless channel and fair scheduling must be employed to provide the proper flow of information in a WSN. In this paper, we develop a channel adaptive MAC protocol with a traffic-aware dynamic power management algorithm for efficient packet scheduling and queuing in a sensor network, with time varying characteristics of the wireless channel also taken into consideration. The proposed protocol calculates a combined weight value based on the channel state and link quality. Then transmission is allowed only for those nodes with weights greater than a minimum quality threshold and nodes attempting to access the wireless medium with a low weight will be allowed to transmit only when their weight becomes high. This results in many poor quality nodes being deprived of transmission for a considerable amount of time. To avoid the buffer overflow and to achieve fairness for the poor quality nodes, we design a Load prediction algorithm. We also design a traffic aware dynamic power management scheme to minimize the energy consumption by continuously turning off the radio interface of all the unnecessary nodes that are not included in the routing path. By Simulation results, we show that our proposed protocol achieves a higher throughput and fairness besides reducing the delay