Advances in scaling deep learning algorithms

Les algorithmes d'apprentissage profond forment un nouvel ensemble de méthodes puissantes pour l'apprentissage automatique. L'idée est de combiner des couches de facteurs latents en hierarchies. Cela requiert souvent un coût computationel plus elevé et augmente aussi le nombre de paramètres du modèle. Ainsi, l'utilisation de ces méthodes sur des problèmes à plus grande échelle demande de réduire leur coût et aussi d'améliorer leur régularisation et leur optimization. Cette thèse adresse cette question sur ces trois perspectives. Nous étudions tout d'abord le problème de réduire le coût de certains algorithmes profonds. Nous proposons deux méthodes pour entrainer des machines de Boltzmann restreintes et des auto-encodeurs débruitants sur des distributions sparses à haute dimension. Ceci est important pour l'application de ces algorithmes pour le traitement de langues naturelles. Ces deux méthodes (Dauphin et al., 2011; Dauphin and Bengio, 2013) utilisent l'échantillonage par importance pour échantilloner l'objectif de ces modèles. Nous observons que cela réduit significativement le temps d'entrainement. L'accéleration atteint 2 ordres de magnitude sur plusieurs bancs d'essai. Deuxièmement, nous introduisont un puissant régularisateur pour les méthodes profondes. Les résultats expérimentaux démontrent qu'un bon régularisateur est crucial pour obtenir de bonnes performances avec des gros réseaux (Hinton et al., 2012). Dans Rifai et al. (2011), nous proposons un nouveau régularisateur qui combine l'apprentissage non-supervisé et la propagation de tangente (Simard et al., 1992). Cette méthode exploite des principes géometriques et permit au moment de la publication d'atteindre des résultats à l'état de l'art. Finalement, nous considérons le problème d'optimiser des surfaces non-convexes à haute dimensionalité comme celle des réseaux de neurones. Tradionellement, l'abondance de minimum locaux était considéré comme la principale difficulté dans ces problèmes. Dans Dauphin et al. (2014a) nous argumentons à partir de résultats en statistique physique, de la théorie des matrices aléatoires, de la théorie des réseaux de neurones et à partir de résultats expérimentaux qu'une difficulté plus profonde provient de la prolifération de points-selle. Dans ce papier nous proposons aussi une nouvelle méthode pour l'optimisation non-convexe.

Nonparametric Estimation of Survivor Function in Bivariate Competing Risk Models’

So far, in the bivariate set up, the analysis of lifetime (failure time) data with multiple causes of failure is done by treating each cause of failure separately. with failures from other causes considered as independent censoring. This approach is unrealistic in many situations. For example, in the analysis of mortality data on married couples one would be interested to compare the hazards for the same cause of death as well as to check whether death due to one cause is more important for the partners’ risk of death from other causes. In reliability analysis. one often has systems with more than one component and many systems. subsystems and components have more than one cause of failure. Design of high-reliability systems generally requires that the individual system components have extremely high reliability even after long periods of time. Knowledge of the failure behaviour of a component can lead to savings in its cost of production and maintenance and. in some cases, to the preservation of human life. For the purpose of improving reliability. it is necessary to identify the cause of failure down to the component level. By treating each cause of failure separately with failures from other causes considered as independent censoring, the analysis of lifetime data would be incomplete. Motivated by this. we introduce a new approach for the analysis of bivariate competing risk data using the bivariate vector hazard rate of Johnson and Kotz (1975).

Computational models for the prediction of severe thunderstorms over east indian region

Thunderstorm is one of the most spectacular weather phenomena in the atmosphere. Many parts over the Indian region experience thunderstorms at higher frequency during pre-monsoon months (March- May), when the atmosphere is highly unstable because of high temperatures prevailing at lower levels. Most dominant feature of the weather during the pre-monsoon season over the eastern Indo-Gangetic plain and northeast India is the outburst of severe local convective storms, commonly known as ‘Nor’wester’ or ‘Kalbaishakhi’. The severe thunderstorms associated with thunder, squall line, lightning and hail cause extensive losses in agriculture, damage to structure and also loss of life. The casualty due to lightning associated with thunderstorms in this region is the highest in the world. The highest numbers of aviation hazards are reported during occurrence of these thunderstorms. In India, 72% of tornadoes are associated with this thunderstorm.

Speaker identification using models for phonemes

Motivation for Speaker recognition work is presented in the first part of the thesis. An exhaustive survey of past work in this field is also presented. A low cost system not including complex computation has been chosen for implementation. Towards achieving this a PC based system is designed and developed. A front end analog to digital convertor (12 bit) is built and interfaced to a PC. Software to control the ADC and to perform various analytical functions including feature vector evaluation is developed. It is shown that a fixed set of phrases incorporating evenly balanced phonemes is aptly suited for the speaker recognition work at hand. A set of phrases are chosen for recognition. Two new methods are adopted for the feature evaluation. Some new measurements involving a symmetry check method for pitch period detection and ACE‘ are used as featured. Arguments are provided to show the need for a new model for speech production. Starting from heuristic, a knowledge based (KB) speech production model is presented. In this model, a KB provides impulses to a voice producing mechanism and constant correction is applied via a feedback path. It is this correction that differs from speaker to speaker. Methods of defining measurable parameters for use as features are described. Algorithms for speaker recognition are developed and implemented. Two methods are presented. The first is based on the model postulated. Here the entropy on the utterance of a phoneme is evaluated. The transitions of voiced regions are used as speaker dependent features. The second method presented uses features found in other works, but evaluated differently. A knock—out scheme is used to provide the weightage values for the selection of features. Results of implementation are presented which show on an average of 80% recognition. It is also shown that if there are long gaps between sessions, the performance deteriorates and is speaker dependent. Cross recognition percentages are also presented and this in the worst case rises to 30% while the best case is 0%. Suggestions for further work are given in the concluding chapter.

Development and Deployment of Wireless Sensor Network in Paddy Fields of Kuttanad

The evolution of wireless sensor network technology has enabled us to develop advanced systems for real time monitoring. In the present scenario wireless sensor networks are increasingly being used for precision agriculture. The advantages of using wireless sensor networks in agriculture are distributed data collection and monitoring, monitor and control of climate, irrigation and nutrient supply. Hence decreasing the cost of production and increasing the efficiency of production. This paper describes the development and deployment of wireless sensor network for crop monitoring in the paddy fields of Kuttanad, a region of Kerala, the southern state of India.

A study on the performance of flexible pavements on mature soil subgrades

The country has witnessed tremendous increase in the vehicle population and increased axle loading pattern during the last decade, leaving its road network overstressed and leading to premature failure. The type of deterioration present in the pavement should be considered for determining whether it has a functional or structural deficiency, so that appropriate overlay type and design can be developed. Structural failure arises from the conditions that adversely affect the load carrying capability of the pavement structure. Inadequate thickness, cracking, distortion and disintegration cause structural deficiency. Functional deficiency arises when the pavement does not provide a smooth riding surface and comfort to the user. This can be due to poor surface friction and texture, hydro planning and splash from wheel path, rutting and excess surface distortion such as potholes, corrugation, faulting, blow up, settlement, heaves etc. Functional condition determines the level of service provided by the facility to its users at a particular time and also the Vehicle Operating Costs (VOC), thus influencing the national economy. Prediction of the pavement deterioration is helpful to assess the remaining effective service life (RSL) of the pavement structure on the basis of reduction in performance levels, and apply various alternative designs and rehabilitation strategies with a long range funding requirement for pavement preservation. In addition, they can predict the impact of treatment on the condition of the sections. The infrastructure prediction models can thus be classified into four groups, namely primary response models, structural performance models, functional performance models and damage models. The factors affecting the deterioration of the roads are very complex in nature and vary from place to place. Hence there is need to have a thorough study of the deterioration mechanism under varied climatic zones and soil conditions before arriving at a definite strategy of road improvement. Realizing the need for a detailed study involving all types of roads in the state with varying traffic and soil conditions, the present study has been attempted. This study attempts to identify the parameters that affect the performance of roads and to develop performance models suitable to Kerala conditions. A critical review of the various factors that contribute to the pavement performance has been presented based on the data collected from selected road stretches and also from five corporations of Kerala. These roads represent the urban conditions as well as National Highways, State Highways and Major District Roads in the sub urban and rural conditions. This research work is a pursuit towards a study of the road condition of Kerala with respect to varying soil, traffic and climatic conditions, periodic performance evaluation of selected roads of representative types and development of distress prediction models for roads of Kerala. In order to achieve this aim, the study is focused into 2 parts. The first part deals with the study of the pavement condition and subgrade soil properties of urban roads distributed in 5 Corporations of Kerala; namely Thiruvananthapuram, Kollam, Kochi, Thrissur and Kozhikode. From selected 44 roads, 68 homogeneous sections were studied. The data collected on the functional and structural condition of the surface include pavement distress in terms of cracks, potholes, rutting, raveling and pothole patching. The structural strength of the pavement was measured as rebound deflection using Benkelman Beam deflection studies. In order to collect the details of the pavement layers and find out the subgrade soil properties, trial pits were dug and the in-situ field density was found using the Sand Replacement Method. Laboratory investigations were carried out to find out the subgrade soil properties, soil classification, Atterberg limits, Optimum Moisture Content, Field Moisture Content and 4 days soaked CBR. The relative compaction in the field was also determined. The traffic details were also collected by conducting traffic volume count survey and axle load survey. From the data thus collected, the strength of the pavement was calculated which is a function of the layer coefficient and thickness and is represented as Structural Number (SN). This was further related to the CBR value of the soil and the Modified Structural Number (MSN) was found out. The condition of the pavement was represented in terms of the Pavement Condition Index (PCI) which is a function of the distress of the surface at the time of the investigation and calculated in the present study using deduct value method developed by U S Army Corps of Engineers. The influence of subgrade soil type and pavement condition on the relationship between MSN and rebound deflection was studied using appropriate plots for predominant types of soil and for classified value of Pavement Condition Index. The relationship will be helpful for practicing engineers to design the overlay thickness required for the pavement, without conducting the BBD test. Regression analysis using SPSS was done with various trials to find out the best fit relationship between the rebound deflection and CBR, and other soil properties for Gravel, Sand, Silt & Clay fractions. The second part of the study deals with periodic performance evaluation of selected road stretches representing National Highway (NH), State Highway (SH) and Major District Road (MDR), located in different geographical conditions and with varying traffic. 8 road sections divided into 15 homogeneous sections were selected for the study and 6 sets of continuous periodic data were collected. The periodic data collected include the functional and structural condition in terms of distress (pothole, pothole patch, cracks, rutting and raveling), skid resistance using a portable skid resistance pendulum, surface unevenness using Bump Integrator, texture depth using sand patch method and rebound deflection using Benkelman Beam. Baseline data of the study stretches were collected as one time data. Pavement history was obtained as secondary data. Pavement drainage characteristics were collected in terms of camber or cross slope using camber board (slope meter) for the carriage way and shoulders, availability of longitudinal side drain, presence of valley, terrain condition, soil moisture content, water table data, High Flood Level, rainfall data, land use and cross slope of the adjoining land. These data were used for finding out the drainage condition of the study stretches. Traffic studies were conducted, including classified volume count and axle load studies. From the field data thus collected, the progression of each parameter was plotted for all the study roads; and validated for their accuracy. Structural Number (SN) and Modified Structural Number (MSN) were calculated for the study stretches. Progression of the deflection, distress, unevenness, skid resistance and macro texture of the study roads were evaluated. Since the deterioration of the pavement is a complex phenomena contributed by all the above factors, pavement deterioration models were developed as non linear regression models, using SPSS with the periodic data collected for all the above road stretches. General models were developed for cracking progression, raveling progression, pothole progression and roughness progression using SPSS. A model for construction quality was also developed. Calibration of HDM–4 pavement deterioration models for local conditions was done using the data for Cracking, Raveling, Pothole and Roughness. Validation was done using the data collected in 2013. The application of HDM-4 to compare different maintenance and rehabilitation options were studied considering the deterioration parameters like cracking, pothole and raveling. The alternatives considered for analysis were base alternative with crack sealing and patching, overlay with 40 mm BC using ordinary bitumen, overlay with 40 mm BC using Natural Rubber Modified Bitumen and an overlay of Ultra Thin White Topping. Economic analysis of these options was done considering the Life Cycle Cost (LCC). The average speed that can be obtained by applying these options were also compared. The results were in favour of Ultra Thin White Topping over flexible pavements. Hence, Design Charts were also plotted for estimation of maximum wheel load stresses for different slab thickness under different soil conditions. The design charts showed the maximum stress for a particular slab thickness and different soil conditions incorporating different k values. These charts can be handy for a design engineer. Fuzzy rule based models developed for site specific conditions were compared with regression models developed using SPSS. The Riding Comfort Index (RCI) was calculated and correlated with unevenness to develop a relationship. Relationships were developed between Skid Number and Macro Texture of the pavement. The effort made through this research work will be helpful to highway engineers in understanding the behaviour of flexible pavements in Kerala conditions and for arriving at suitable maintenance and rehabilitation strategies. Key Words: Flexible Pavements – Performance Evaluation – Urban Roads – NH – SH and other roads – Performance Models – Deflection – Riding Comfort Index – Skid Resistance – Texture Depth – Unevenness – Ultra Thin White Topping

Coordinating Inventory Control and Pricing Strategies with Random Demand and Fixed Ordering Cost: the Infinite Horizon Case

We analyze an infinite horizon, single product, periodic review model in which pricing and production/inventory decisions are made simultaneously. Demands in different periods are identically distributed random variables that are independent of each other and their distributions depend on the product price. Pricing and ordering decisions are made at the beginning of each period and all shortages are backlogged. Ordering cost includes both a fixed cost and a variable cost proportional to the amount ordered. The objective is to maximize expected discounted, or expected average profit over the infinite planning horizon. We show that a stationary (s,S,p) policy is optimal for both the discounted and average profit models with general demand functions. In such a policy, the period inventory is managed based on the classical (s,S) policy and price is determined based on the inventory position at the beginning of each period.

Social network measures for "nosduocentered" networks, their predictive power on performance

Our purpose in this article is to define a network structure which is based on two egos instead of the egocentered (one ego) or the complete network (n egos). We describe the characteristics and properties for this kind of network which we call “nosduocentered network”, comparing it with complete and egocentered networks. The key point for this kind of network is that relations exist between the two main egos and all alters, but relations among others are not observed. After that, we use new social network measures adapted to the nosduocentered network, some of which are based on measures for complete networks such as degree, betweenness, closeness centrality or density, while some others are tailormade for nosduocentered networks. We specify three regression models to predict research performance of PhD students based on these social network measures for different networks such as advice, collaboration, emotional support and trust. Data used are from Slovenian PhD students and their s

Static Pricing for a Network Service Provider

This article studies the static pricing problem of a network service provider who has a fixed capacity and faces different types of customers (classes). Each type of customers can have its own capacity constraint but it is assumed that all classes have the same resource requirement. The provider must decide a static price for each class. The customer types are characterized by their arrival process, with a price-dependant arrival rate, and the random time they remain in the system. Many real-life situations could fit in this framework, for example an Internet provider or a call center, but originally this problem was thought for a company that sells phone-cards and needs to set the price-per-minute for each destination. Our goal is to characterize the optimal static prices in order to maximize the provider's revenue. We note that the model here presented, with some slight modifications and additional assumptions can be used in those cases when the objective is to maximize social welfare.

A quality-based cost distribution charging scheme for QoS multicast networks

This paper presents a new charging scheme for cost distribution along a point-to-multipoint connection when destination nodes are responsible for the cost. The scheme focus on QoS considerations and a complete range of choices is presented. These choices go from a safe scheme for the network operator to a fair scheme to the customer. The in-between cases are also covered. Specific and general problems, like the incidence of users disconnecting dynamically is also discussed. The aim of this scheme is to encourage the users to disperse the resource demand instead of having a large number of direct connections to the source of the data, which would result in a higher than necessary bandwidth use from the source. This would benefit the overall performance of the network. The implementation of this task must balance between the necessity to offer a competitive service and the risk of not recovering such service cost for the network operator. Throughout this paper reference to multicast charging is made without making any reference to any specific category of service. The proposed scheme is also evaluated with the criteria set proposed in the European ATM charging project CANCAN

Operational Cost Reduction in WDM Networks using Lighttours

Wavelength division multiplexing (WDM) networks have been adopted as a near-future solution for the broadband Internet. In previous work we proposed a new architecture, named enhanced grooming (G+), that extends the capabilities of traditional optical routes (lightpaths). In this paper, we compare the operational expenditures incurred by routing a set of demands using lightpaths with that of lighttours. The comparison is done by solving an integer linear programming (ILP) problem based on a path formulation. Results show that, under the assumption of single-hop routing, almost 15% of the operational cost can be reduced with our architecture. In multi-hop routing the operation cost is reduced in 7.1% and at the same time the ratio of operational cost to number of optical-electro-optical conversions is reduced for our architecture. This means that ISPs could provide the same satisfaction in terms of delay to the end-user with a lower investment in the network architecture

An ATM distributed simulator for network management research

Due to the high cost of a large ATM network working up to full strength to apply our ideas about network management, i.e., dynamic virtual path (VP) management and fault restoration, we developed a distributed simulation platform for performing our experiments. This platform also had to be capable of other sorts of tests, such as connection admission control (CAC) algorithms, routing algorithms, and accounting and charging methods. The platform was posed as a very simple, event-oriented and scalable simulation. The main goal was the simulation of a working ATM backbone network with a potentially large number of nodes (hundreds). As research into control algorithms and low-level, or rather cell-level methods, was beyond the scope of this study, the simulation took place at a connection level, i.e., there was no real traffic of cells. The simulated network behaved like a real network accepting and rejecting SNMP ones, or experimental tools using the API node

All-Optical Label Stacking: Easing the Trade-offs Between Routing and Architecture Cost in All-Optical Packet Switching

All-optical label swapping (AOLS) forms a key technology towards the implementation of all-optical packet switching nodes (AOPS) for the future optical Internet. The capital expenditures of the deployment of AOLS increases with the size of the label spaces (i.e. the number of used labels), since a special optical device is needed for each recognized label on every node. Label space sizes are affected by the way in which demands are routed. For instance, while shortest-path routing leads to the usage of fewer labels but high link utilization, minimum interference routing leads to the opposite. This paper studies all-optical label stacking (AOLStack), which is an extension of the AOLS architecture. AOLStack aims at reducing label spaces while easing the compromise with link utilization. In this paper, an integer lineal program is proposed with the objective of analyzing the softening of the aforementioned trade-off due to AOLStack. Furthermore, a heuristic aiming at finding good solutions in polynomial-time is proposed as well. Simulation results show that AOLStack either a) reduces the label spaces with a low increase in the link utilization or, similarly, b) uses better the residual bandwidth to decrease the number of labels even more

A hybrid method for sag source location in power network

The work presented in this paper belongs to the power quality knowledge area and deals with the voltage sags in power transmission and distribution systems. Propagating throughout the power network, voltage sags can cause plenty of problems for domestic and industrial loads that can financially cost a lot. To impose penalties to responsible party and to improve monitoring and mitigation strategies, sags must be located in the power network. With such a worthwhile objective, this paper comes up with a new method for associating a sag waveform with its origin in transmission and distribution networks. It solves this problem through developing hybrid methods which hire multiway principal component analysis (MPCA) as a dimension reduction tool. MPCA reexpresses sag waveforms in a new subspace just in a few scores. We train some well-known classifiers with these scores and exploit them for classification of future sags. The capabilities of the proposed method for dimension reduction and classification are examined using the real data gathered from three substations in Catalonia, Spain. The obtained classification rates certify the goodness and powerfulness of the developed hybrid methods as brand-new tools for sag classification

La adopción de un enfoque de trabajo en red: implicaciones en torno al proceso estratégico de las organizaciones

En el mundo actual el enfoque del proceso estratégico es una variable que cada vez va tomando un mayor peso e incidencia en las decisiones de los directores de empresa, sin embargo, poco se ha estudiado sobre esta variable al ser adoptado por los nuevos modelos de cooperación empresarial en las organizaciones. Es por esta razón que esta investigación busca esclarecer la aplicabilidad de las metodologías y herramientas utilizadas en la planeación estratégica de una empresa en una red, además de servir como un primer acercamiento a los aspectos más importantes en la planeación estratégica de una red.