Deep learning of representations and its application to computer vision

L’objectif de cette thèse par articles est de présenter modestement quelques étapes du parcours qui mènera (on espère) à une solution générale du problème de l’intelligence artificielle. Cette thèse contient quatre articles qui présentent chacun une différente nouvelle méthode d’inférence perceptive en utilisant l’apprentissage machine et, plus particulièrement, les réseaux neuronaux profonds. Chacun de ces documents met en évidence l’utilité de sa méthode proposée dans le cadre d’une tâche de vision par ordinateur. Ces méthodes sont applicables dans un contexte plus général, et dans certains cas elles on tété appliquées ailleurs, mais ceci ne sera pas abordé dans le contexte de cette de thèse. Dans le premier article, nous présentons deux nouveaux algorithmes d’inférence variationelle pour le modèle génératif d’images appelé codage parcimonieux “spike- and-slab” (CPSS). Ces méthodes d’inférence plus rapides nous permettent d’utiliser des modèles CPSS de tailles beaucoup plus grandes qu’auparavant. Nous démontrons qu’elles sont meilleures pour extraire des détecteur de caractéristiques quand très peu d’exemples étiquetés sont disponibles pour l’entraînement. Partant d’un modèle CPSS, nous construisons ensuite une architecture profonde, la machine de Boltzmann profonde partiellement dirigée (MBP-PD). Ce modèle a été conçu de manière à simplifier d’entraînement des machines de Boltzmann profondes qui nécessitent normalement une phase de pré-entraînement glouton pour chaque couche. Ce problème est réglé dans une certaine mesure, mais le coût d’inférence dans le nouveau modèle est relativement trop élevé pour permettre de l’utiliser de manière pratique. Dans le deuxième article, nous revenons au problème d’entraînement joint de machines de Boltzmann profondes. Cette fois, au lieu de changer de famille de modèles, nous introduisons un nouveau critère d’entraînement qui donne naissance aux machines de Boltzmann profondes à multiples prédictions (MBP-MP). Les MBP-MP sont entraînables en une seule étape et ont un meilleur taux de succès en classification que les MBP classiques. Elles s’entraînent aussi avec des méthodes variationelles standard au lieu de nécessiter un classificateur discriminant pour obtenir un bon taux de succès en classification. Par contre, un des inconvénients de tels modèles est leur incapacité de générer deséchantillons, mais ceci n’est pas trop grave puisque la performance de classification des machines de Boltzmann profondes n’est plus une priorité étant donné les dernières avancées en apprentissage supervisé. Malgré cela, les MBP-MP demeurent intéressantes parce qu’elles sont capable d’accomplir certaines tâches que des modèles purement supervisés ne peuvent pas faire, telles que celle de classifier des données incomplètes ou encore celle de combler intelligemment l’information manquante dans ces données incomplètes. Le travail présenté dans cette thèse s’est déroulé au milieu d’une période de transformations importantes du domaine de l’apprentissage à réseaux neuronaux profonds qui a été déclenchée par la découverte de l’algorithme de “dropout” par Geoffrey Hinton. Dropout rend possible un entraînement purement supervisé d’architectures de propagation unidirectionnel sans être exposé au danger de sur- entraînement. Le troisième article présenté dans cette thèse introduit une nouvelle fonction d’activation spécialement con ̧cue pour aller avec l’algorithme de Dropout. Cette fonction d’activation, appelée maxout, permet l’utilisation de aggrégation multi-canal dans un contexte d’apprentissage purement supervisé. Nous démontrons comment plusieurs tâches de reconnaissance d’objets sont mieux accomplies par l’utilisation de maxout. Pour terminer, sont présentons un vrai cas d’utilisation dans l’industrie pour la transcription d’adresses de maisons à plusieurs chiffres. En combinant maxout avec une nouvelle sorte de couche de sortie pour des réseaux neuronaux de convolution, nous démontrons qu’il est possible d’atteindre un taux de succès comparable à celui des humains sur un ensemble de données coriace constitué de photos prises par les voitures de Google. Ce système a été déployé avec succès chez Google pour lire environ cent million d’adresses de maisons.

Articulated Model Registration of MRI/X-Ray Spine Data

This paper presents a method based on articulated models for the registration of spine data extracted from multimodal medical images of patients with scoliosis. With the ultimate aim being the development of a complete geometrical model of the torso of a scoliotic patient, this work presents a method for the registration of vertebral column data using 3D magnetic resonance images (MRI) acquired in prone position and X-ray data acquired in standing position for five patients with scoliosis. The 3D shape of the vertebrae is estimated from both image modalities for each patient, and an articulated model is used in order to calculate intervertebral transformations required in order to align the vertebrae between both postures. Euclidean distances between anatomical landmarks are calculated in order to assess multimodal registration error. Results show a decrease in the Euclidean distance using the proposed method compared to rigid registration and more physically realistic vertebrae deformations compared to thin-plate-spline (TPS) registration thus improving alignment.

Studies Using Computer Simulation Models to Solve Problems Related to Logistics Terminals

This thesis deals with the use of simulation as a problem-solving tool to solve a few logistic system related problems. More specifically it relates to studies on transport terminals. Transport terminals are key elements in the supply chains of industrial systems. One of the problems related to use of simulation is that of the multiplicity of models needed to study different problems. There is a need for development of methodologies related to conceptual modelling which will help reduce the number of models needed. Three different logistic terminal systems Viz. a railway yard, container terminal of apart and airport terminal were selected as cases for this study. The standard methodology for simulation development consisting of system study and data collection, conceptual model design, detailed model design and development, model verification and validation, experimentation, and analysis of results, reporting of finding were carried out. We found that models could be classified into tightly pre-scheduled, moderately pre-scheduled and unscheduled systems. Three types simulation models( called TYPE 1, TYPE 2 and TYPE 3) of various terminal operations were developed in the simulation package Extend. All models were of the type discrete-event simulation. Simulation models were successfully used to help solve strategic, tactical and operational problems related to three important logistic terminals as set in our objectives. From the point of contribution to conceptual modelling we have demonstrated that clubbing problems into operational, tactical and strategic and matching them with tightly pre-scheduled, moderately pre-scheduled and unscheduled systems is a good workable approach which reduces the number of models needed to study different terminal related problems.

An Alternative Approach To Computer System Security Monitoring And Enhancement Through System Call Sequence Analysis

Modern computer systems are plagued with stability and security problems: applications lose data, web servers are hacked, and systems crash under heavy load. Many of these problems or anomalies arise from rare program behavior caused by attacks or errors. A substantial percentage of the web-based attacks are due to buffer overflows. Many methods have been devised to detect and prevent anomalous situations that arise from buffer overflows. The current state-of-art of anomaly detection systems is relatively primitive and mainly depend on static code checking to take care of buffer overflow attacks. For protection, Stack Guards and I-leap Guards are also used in wide varieties.This dissertation proposes an anomaly detection system, based on frequencies of system calls in the system call trace. System call traces represented as frequency sequences are profiled using sequence sets. A sequence set is identified by the starting sequence and frequencies of specific system calls. The deviations of the current input sequence from the corresponding normal profile in the frequency pattern of system calls is computed and expressed as an anomaly score. A simple Bayesian model is used for an accurate detection.Experimental results are reported which show that frequency of system calls represented using sequence sets, captures the normal behavior of programs under normal conditions of usage. This captured behavior allows the system to detect anomalies with a low rate of false positives. Data are presented which show that Bayesian Network on frequency variations responds effectively to induced buffer overflows. It can also help administrators to detect deviations in program flow introduced due to errors.

Management of Traffic Control — A Computer Oriented Approach

Traffic Management system (TMS) comprises four major sub systems: The Network Database Management system for information to the passengers, Transit Facility Management System for service, planning, and scheduling vehicle and crews, Congestion Management System for traffic forecasting and planning, Safety Management System concerned with safety aspects of passengers and Environment. This work has opened a rather wide frame work of model structures for application on traffic. The facets of these theories are so wide that it seems impossible to present all necessary models in this work. However it could be deduced from the study that the best Traffic Management System is that whichis realistic in all aspects is easy to understand is easy to apply As it is practically difficult to device an ideal fool—proof model, the attempt here has been to make some progress-in that direction.

A Comparative Study of Convective Parameterization Schemes in WRF-NMM Model

Severe local storms, including tornadoes, damaging hail and wind gusts, frequently occur over the eastern and northeastern states of India during the pre-monsoon season (March-May). Forecasting thunderstorms is one of the most difficult tasks in weather prediction, due to their rather small spatial and temporal extension and the inherent non-linearity of their dynamics and physics. In this paper, sensitivity experiments are conducted with the WRF-NMM model to test the impact of convective parameterization schemes on simulating severe thunderstorms that occurred over Kolkata on 20 May 2006 and 21 May 2007 and validated the model results with observation. In addition, a simulation without convective parameterization scheme was performed for each case to determine if the model could simulate the convection explicitly. A statistical analysis based on mean absolute error, root mean square error and correlation coefficient is performed for comparisons between the simulated and observed data with different convective schemes. This study shows that the prediction of thunderstorm affected parameters is sensitive to convective schemes. The Grell-Devenyi cloud ensemble convective scheme is well simulated the thunderstorm activities in terms of time, intensity and the region of occurrence of the events as compared to other convective schemes and also explicit scheme

A model of Distributed Ayurvedic Diagnosis and Therapy System for Hridroga Using Agents

Diagnosis of Hridroga (cardiac disorders) in Ayurveda requires the combination of many different types of data, including personal details, patient symptoms, patient histories, general examination results, Ashtavidha pareeksha results etc. Computer-assisted decision support systems must be able to combine these data types into a seamless system. Intelligent agents, an approach that has been used chiefly in business applications, is used in medical diagnosis in this case. This paper is about a multi-agent system named “Distributed Ayurvedic Diagnosis and Therapy System for Hridroga using Agents” (DADTSHUA). It describes the architecture of the DADTSHUA model .This system is using mobile agents and ontology for passing data through the network. Due to this, transport delay can be minimized. It is a system which will be very helpful for the beginning physicians to eliminate his ambiguity in diagnosis and therapy. The system is implemented using Java Agent DEvelopment framework (JADE), which is a java-complaint mobile agent platform from TILab.

Genetic Programming meets Model-Driven Development

Genetic programming is known to provide good solutions for many problems like the evolution of network protocols and distributed algorithms. In such cases it is most likely a hardwired module of a design framework that assists the engineer to optimize specific aspects of the system to be developed. It provides its results in a fixed format through an internal interface. In this paper we show how the utility of genetic programming can be increased remarkably by isolating it as a component and integrating it into the model-driven software development process. Our genetic programming framework produces XMI-encoded UML models that can easily be loaded into widely available modeling tools which in turn posses code generation as well as additional analysis and test capabilities. We use the evolution of a distributed election algorithm as an example to illustrate how genetic programming can be combined with model-driven development. This example clearly illustrates the advantages of our approach – the generation of source code in different programming languages.

Numerical aspects in the data model of conceptual information systems

While most data analysis and decision support tools use numerical aspects of the data, Conceptual Information Systems focus on their conceptual structure. This paper discusses how both approaches can be combined.

A finite state model for on-line analytical processing in triadic contexts

About ten years ago, triadic contexts were presented by Lehmann and Wille as an extension of Formal Concept Analysis. However, they have rarely been used up to now, which may be due to the rather complex structure of the resulting diagrams. In this paper, we go one step back and discuss how traditional line diagrams of standard (dyadic) concept lattices can be used for exploring and navigating triadic data. Our approach is inspired by the slice & dice paradigm of On-Line-Analytical Processing (OLAP). We recall the basic ideas of OLAP, and show how they may be transferred to triadic contexts. For modeling the navigation patterns a user might follow, we use the formalisms of finite state machines. In order to present the benefits of our model, we show how it can be used for navigating the IT Baseline Protection Manual of the German Federal Office for Information Security.

Development of an ArcGIS extension to model urban climate factors

The possibility to develop automatically running models which can capture some of the most important factors driving the urban climate would be very useful for many planning aspects. With the help of these modulated climate data, the creation of the typically used “Urban Climate Maps” (UCM) will be accelerated and facilitated. This work describes the development of a special ArcGIS software extension, along with two support databases to achieve this functionality. At the present time, lacking comparability between different UCMs and imprecise planning advices going along with the significant technical problems of manually creating conventional maps are central issues. Also inflexibility and static behaviour are reducing the maps’ practicality. From experi-ence, planning processes are formed more productively, namely to implant new planning parameters directly via the existing work surface to map the impact of the data change immediately, if pos-sible. In addition to the direct climate figures, information of other planning areas (like regional characteristics / developments etc.) have to be taken into account to create the UCM as well. Taking all these requirements into consideration, an automated calculation process of urban climate impact parameters will serve to increase the creation of homogenous UCMs efficiently.

Structural Dynamics Model of a Cartesian Robot

Methods are developed for predicting vibration response characteristics of systems which change configuration during operation. A cartesian robot, an example of such a position-dependent system, served as a test case for these methods and was studied in detail. The chosen system model was formulated using the technique of Component Mode Synthesis (CMS). The model assumes that he system is slowly varying, and connects the carriages to each other and to the robot structure at the slowly varying connection points. The modal data required for each component is obtained experimentally in order to get a realistic model. The analysis results in prediction of vibrations that are produced by the inertia forces as well as gravity and friction forces which arise when the robot carriages move with some prescribed motion. Computer simulations and experimental determinations are conducted in order to calculate the vibrations at the robot end-effector. Comparisons are shown to validate the model in two ways: for fixed configuration the mode shapes and natural frequencies are examined, and then for changing configuration the residual vibration at the end of the mode is evaluated. A preliminary study was done on a geometrically nonlinear system which also has position-dependency. The system consisted of a flexible four-bar linkage with elastic input and output shafts. The behavior of the rocker-beam is analyzed for different boundary conditions to show how some limiting cases are obtained. A dimensional analysis leads to an evaluation of the consequences of dynamic similarity on the resulting vibration.

Procedures as a Representation for Data in a Computer Program for Understanding Natural Language

This paper describes a system for the computer understanding of English. The system answers questions, executes commands, and accepts information in normal English dialog. It uses semantic information and context to understand discourse and to disambiguate sentences. It combines a complete syntactic analysis of each sentence with a "heuristic understander" which uses different kinds of information about a sentence, other parts of the discourse, and general information about the world in deciding what the sentence means. It is based on the belief that a computer cannot deal reasonably with language unless it can "understand" the subject it is discussing. The program is given a detailed model of the knowledge needed by a simple robot having only a hand and an eye. We can give it instructions to manipulate toy objects, interrogate it about the scene, and give it information it will use in deduction. In addition to knowing the properties of toy objects, the program has a simple model of its own mentality. It can remember and discuss its plans and actions as well as carry them out. It enters into a dialog with a person, responding to English sentences with actions and English replies, and asking for clarification when its heuristic programs cannot understand a sentence through use of context and physical knowledge.

ADAM: A Decentralized Parallel Computer Architecture Featuring Fast Thread and Data Migration and a Uniform Hardware Abstraction

The furious pace of Moore's Law is driving computer architecture into a realm where the the speed of light is the dominant factor in system latencies. The number of clock cycles to span a chip are increasing, while the number of bits that can be accessed within a clock cycle is decreasing. Hence, it is becoming more difficult to hide latency. One alternative solution is to reduce latency by migrating threads and data, but the overhead of existing implementations has previously made migration an unserviceable solution so far. I present an architecture, implementation, and mechanisms that reduces the overhead of migration to the point where migration is a viable supplement to other latency hiding mechanisms, such as multithreading. The architecture is abstract, and presents programmers with a simple, uniform fine-grained multithreaded parallel programming model with implicit memory management. In other words, the spatial nature and implementation details (such as the number of processors) of a parallel machine are entirely hidden from the programmer. Compiler writers are encouraged to devise programming languages for the machine that guide a programmer to express their ideas in terms of objects, since objects exhibit an inherent physical locality of data and code. The machine implementation can then leverage this locality to automatically distribute data and threads across the physical machine by using a set of high performance migration mechanisms. An implementation of this architecture could migrate a null thread in 66 cycles -- over a factor of 1000 improvement over previous work. Performance also scales well; the time required to move a typical thread is only 4 to 5 times that of a null thread. Data migration performance is similar, and scales linearly with data block size. Since the performance of the migration mechanism is on par with that of an L2 cache, the implementation simulated in my work has no data caches and relies instead on multithreading and the migration mechanism to hide and reduce access latencies.

Model-Based Matching by Linear Combinations of Prototypes

We describe a method for modeling object classes (such as faces) using 2D example images and an algorithm for matching a model to a novel image. The object class models are "learned'' from example images that we call prototypes. In addition to the images, the pixelwise correspondences between a reference prototype and each of the other prototypes must also be provided. Thus a model consists of a linear combination of prototypical shapes and textures. A stochastic gradient descent algorithm is used to match a model to a novel image by minimizing the error between the model and the novel image. Example models are shown as well as example matches to novel images. The robustness of the matching algorithm is also evaluated. The technique can be used for a number of applications including the computation of correspondence between novel images of a certain known class, object recognition, image synthesis and image compression.