Understanding deep architectures and the effect of unsupervised pre-training

Cette thèse porte sur une classe d'algorithmes d'apprentissage appelés architectures profondes. Il existe des résultats qui indiquent que les représentations peu profondes et locales ne sont pas suffisantes pour la modélisation des fonctions comportant plusieurs facteurs de variation. Nous sommes particulièrement intéressés par ce genre de données car nous espérons qu'un agent intelligent sera en mesure d'apprendre à les modéliser automatiquement; l'hypothèse est que les architectures profondes sont mieux adaptées pour les modéliser. Les travaux de Hinton (2006) furent une véritable percée, car l'idée d'utiliser un algorithme d'apprentissage non-supervisé, les machines de Boltzmann restreintes, pour l'initialisation des poids d'un réseau de neurones supervisé a été cruciale pour entraîner l'architecture profonde la plus populaire, soit les réseaux de neurones artificiels avec des poids totalement connectés. Cette idée a été reprise et reproduite avec succès dans plusieurs contextes et avec une variété de modèles. Dans le cadre de cette thèse, nous considérons les architectures profondes comme des biais inductifs. Ces biais sont représentés non seulement par les modèles eux-mêmes, mais aussi par les méthodes d'entraînement qui sont souvent utilisés en conjonction avec ceux-ci. Nous désirons définir les raisons pour lesquelles cette classe de fonctions généralise bien, les situations auxquelles ces fonctions pourront être appliquées, ainsi que les descriptions qualitatives de telles fonctions. L'objectif de cette thèse est d'obtenir une meilleure compréhension du succès des architectures profondes. Dans le premier article, nous testons la concordance entre nos intuitions---que les réseaux profonds sont nécessaires pour mieux apprendre avec des données comportant plusieurs facteurs de variation---et les résultats empiriques. Le second article est une étude approfondie de la question: pourquoi l'apprentissage non-supervisé aide à mieux généraliser dans un réseau profond? Nous explorons et évaluons plusieurs hypothèses tentant d'élucider le fonctionnement de ces modèles. Finalement, le troisième article cherche à définir de façon qualitative les fonctions modélisées par un réseau profond. Ces visualisations facilitent l'interprétation des représentations et invariances modélisées par une architecture profonde.

Dynamic and Thermodynamic Structure of the Monsoon Boundary Layer over India

The objective of the study is to examine the dynamic and thermodynamic structure and the variations that occur in the surface layer during the pre-monsoon, onset and post-monsoon periods over the Indian region. The variations caused during the occurrence of micro and mesoscale systems, structure and variation in the marine boundary layer over the Indian region is also investigated. The drag coefficient computed indirectly also shows variation during various seasons. The thermodynamic structure of the atmosphere shows variation during the various seasons. The onset monsoon causes lowering of the Lifting Condensation Levels. The outcome of the study is expected to provide a better understanding of the structure and variations in the boundary layer over India, which is useful for many applications especially for numerical modeling studies.

Characteristic Study of the Boundary Layer Parameters over the Arabian Sea and the Bay of Bengal Using the QuikSCAT Dataset

The marine atmospheric boundary layer (MABL) plays a vital role in the transport of momentum and heat from the surface of the ocean into the atmosphere. A detailed study on the MABL characteristics was carried out using high-resolution surface-wind data as measured by the QuikSCAT (Quick scatterometer) satellite. Spatial variations in the surface wind, frictional velocity, roughness parameter and drag coe±cient for the di®erent seasons were studied. The surface wind was strong during the southwest monsoon season due to the modulation induced by the Low Level Jetstream. The drag coe±cient was larger during this season, due to the strong winds and was lower during the winter months. The spatial variations in the frictional velocity over the seas was small during the post-monsoon season (»0.2 m s¡1). The maximum spatial variation in the frictional velocity was found over the south Arabian Sea (0.3 to 0.5 m s¡1) during the southwest monsoon period, followed by the pre-monsoon over the Bay of Bengal (0.1 to 0.25 m s¡1). The mean wind-stress curl during the winter was positive over the equatorial region, with a maximum value of 1.5£10¡7 N m¡3, but on either side of the equatorial belt, a negative wind-stress curl dominated. The area average of the frictional velocity and drag coe±cient over the Arabian Sea and Bay of Bengal were also studied. The values of frictional velocity shows a variability that is similar to the intraseasonal oscillation (ISO) and this was con¯rmed via wavelet analysis. In the case of the drag coe±cient, the prominent oscillations were ISO and quasi-biweekly mode (QBM). The interrelationship between the drag coe±cient and the frictional velocity with wind speed in both the Arabian Sea and the Bay of Bengal was also studied.

Chlorine tolerant, multilayer reverse-somosis membranes with high permeate flux and high salt rejection

A new class of high molecular weight polyethersulfone ionomers is described in which the ionic content can be varied, at will, over a very wide and fully-controllable range. A novel type of coating process enables these materials to be deposited from alcohol-type solvents as cohesive but very thin (50 – 250 nm) films on porous support-membranes, giving high-flux membranes (3.3 – 5.0 L m-2 h-1 bar-1) with very good, though not outstanding salt rejection (typically 92 - 96%). A secondary layer, of formaldehyde-cross-linked polyvinyl alcohol, can be deposited from aqueous solution on the surface of the ionomer membrane, and this layer increases salt rejection to greater than 99% without serious loss of water permeability. The final multi-layer membrane shows excellent chlorine tolerance in reverse-osmosis operation.

Enhancing consumer liking of low salt tomato soup over repeated exposure by herb and spice seasonings

There is strong evidence for the link between high dietary sodium and increased risk of cardiovascular disease which drives the need to reduce salt content in foods. In this study, herb and spice blends were used to enhance consumer acceptability of a low salt tomato soup (0.26% w/w). Subjects (n=148) scored their liking of tomato soup samples over five consecutive days. The first and last days were pre-and post-exposure visits where all participants rated three tomato soup samples; standard, low salt and low salt with added herbs and spices. The middle 3 days were the repeated exposure phase where participants were divided into three balanced groups; consuming the standard soup, the low salt soup, or the low salt soup with added herbs and spices. Reducing salt in the tomato soup led to a significant decline in consumer acceptability, and incorporating herbs and spices did not lead to an immediate enhancement in liking. However, inclusion of herbs and spices enhanced the perception of the salty taste of the low salt soup to the same level as the standard. Repeated exposure to the herbs and spice-modified soup led to a significant increase in the overall liking and liking of flavour, texture and aftertaste of the soup, whereas no changes in liking were observed for the standard and low salt tomato soups over repeated exposure. Moreover, a positive trend in increasing the post-exposure liking of the herbs and spices soup was observed. The findings suggest that the use of herbs and spices is a useful approach to reduce salt content in foods; however, herbs and spices should be chosen carefully to complement the food as large contrasts in flavour can polarise consumer liking.

Coordinated Cluster and ground-based instrument observations of transient changes in the magnetopause boundary layer during an interval of predominantly northward IMF: relation to reconnection pulses and FTE signatures

We study a series of transient entries into the low-latitude boundary layer (LLBL) of all four Cluster spacecraft during an outbound pass through the mid-afternoon magnetopause ([X(GSM), Y(GSM), Z(GSM)] approximate to [2, 7, 9] R(E)). The events take place during an interval of northward IMF, as seen in the data from the ACE satellite and lagged by a propagation delay of 75 min that is well-defined by two separate studies: (1) the magnetospheric variations prior to the northward turning (Lockwood et al., 2001, this issue) and (2) the field clock angle seen by Cluster after it had emerged into the magnetosheath (Opgenoorth et al., 2001, this issue). With an additional lag of 16.5 min, the transient LLBL events cor-relate well with swings of the IMF clock angle (in GSM) to near 90degrees. Most of this additional lag is explained by ground-based observations, which reveal signatures of transient reconnection in the pre-noon sector that then take 10-15 min to propagate eastward to 15 MLT, where they are observed by Cluster. The eastward phase speed of these signatures agrees very well with the motion deduced by the cross-correlation of the signatures seen on the four Cluster spacecraft. The evidence that these events are reconnection pulses includes: transient erosion of the noon 630 nm (cusp/cleft) aurora to lower latitudes; transient and travelling enhancements of the flow into the polar cap, imaged by the AMIE technique; and poleward-moving events moving into the polar cap, seen by the EISCAT Svalbard Radar (ESR). A pass of the DMSP-F15 satellite reveals that the open field lines near noon have been opened for some time: the more recently opened field lines were found closer to dusk where the flow transient and the poleward-moving event intersected the satellite pass. The events at Cluster have ion and electron characteristics predicted and observed by Lockwood and Hapgood (1998) for a Flux Transfer Event (FTE), with allowance for magnetospheric ion reflection at Alfvenic disturbances in the magnetopause reconnection layer. Like FTEs, the events are about 1 R(E) in their direction of motion and show a rise in the magnetic field strength, but unlike FTEs, in general, they show no pressure excess in their core and hence, no characteristic bipolar signature in the boundary-normal component. However, most of the events were observed when the magnetic field was southward, i.e. on the edge of the interior magnetic cusp, or when the field was parallel to the magnetic equatorial plane. Only when the satellite begins to emerge from the exterior boundary (when the field was northward), do the events start to show a pressure excess in their core and the consequent bipolar signature. We identify the events as the first observations of FTEs at middle altitudes.

Syncytiotrophoblast extracellular vesicles from pre-eclampsia placentas differentially affect platelet function

Pre-eclampsia (PE) complicates around 3% of all pregnancies and is one of the most common causes of maternal mortality worldwide. The pathophysiology of PE remains unclear however its underlying cause originates from the placenta and manifests as raised blood pressure, proteinuria, vascular or systemic inflammation and hypercoagulation in the mother. Women who develop PE are also at significantly higher risk of subsequently developing cardiovascular (CV) disease. In PE, the failing endoplasmic reticulum, oxidative and inflammatory stressed syncytiotrophoblast layer of the placenta sheds increased numbers of syncytiotrophoblast extracellular vesicles (STBEV) into the maternal circulation. Platelet reactivity, size and concentration are also known to be altered in some women who develop PE, although the underlying reasons for this have not been determined. In this study we show that STBEV from disease free placenta isolated ex vivo by dual placental perfusion associate rapidly with platelets. We provide evidence that STBEV isolated from normal placentas cause platelet activation and that this is increased with STBEV from PE pregnancies. Furthermore, treatment of platelets with aspirin, currently prescribed for women at high risk of PE to reduce platelet aggregation, also inhibits STBEV-induced reversible aggregation of washed platelets. Increased platelet reactivity as a result of exposure to PE placenta derived STBEVs correlates with increased thrombotic risk associated with PE. These observations establish a possible direct link between the clotting disturbances of PE and dysfunction of the placenta, as well as the known increased risk of thromboembolism associated with this condition.

Controlled fabrication of gold nanoparticles biomediated by glucose oxidase immobilized on chitosan layer-by-layer films

The control of size and shape of metallic nanoparticles is a fundamental goal in nanochemistry, and crucial for applications exploiting nanoscale properties of materials. We present here an approach to the synthesis of gold nanoparticles mediated by glucose oxidase (GOD) immobilized on solid substrates using the Layer-by-Layer (LbL) technique. The LbL films contained four alternated layers of chitosan and poly(styrene sulfonate) (PSS), with GOD in the uppermost bilayer adsorbed on a fifth chitosan layer: (chitosan/PSS)(4)/(chitosan/GOD). The films were inserted into a solution containing gold salt and glucose, at various pHs. Optimum conditions were achieved at pH 9, producing gold nanoparticles of ca. 30 nm according to transmission electron microscopy. A comparative study with the enzyme in solution demonstrated that the synthesis of gold nanoparticles is more efficient using immobilized GOD. (C) 2009 Elsevier B.V. All rights reserved.

Modeling Adsorption Processes of Poly-p-phenylenevinylene Precursor and Sodium Acid Dodecylbenzenesulfonate onto Layer-by-Layer Films Using a Langmuir-type Metastable Equilibrium Model

The adsorption kinetics curves of poly(xylylidene tetrahydrothiophenium chloride) (PTHT), a poly-p-phenylenevinylene (PPV) precursor, and the sodium salt of dodecylbenzene sulfonic acid (DBS), onto (PTHT/DBS)(n) layer-by-layer (LBL) films were characterized by means of UV-vis spectroscopy. The amount of PTHT/DBS and PTHT adsorbed on each layer was shown to be practically independent of adsorption time. A Langmuir-type metastable equilibrium model was used to adjust the adsorption isotherms data and to estimate adsorption/desorption coefficients ratios, k = k(ads)/k(des), values of 2 x 10(5) and 4 x 10(6) for PTHT and PTHT/DBS layers, respectively. The desorption coefficient has been estimated, using literature values for poly(o-methoxyaniline) desorption coefficient, as was found to be in the range of 10(-9) to 10(-6) s(-1), indicating that quasi equilibrium is rapidly attained.

Immobilization of liposomes in nanostructured layer-by-layer films containing dendrimers

Artificial vesicles or liposomes composed of lipid bilayers have been widely exploited as building blocks for artificial membranes, in attempts to mimic membrane interaction with drugs and proteins and to investigate drug delivery processes. In this study we report on the immobilization of liposomes of 1,2-dipalmitoyi-sn-Glycero-3-[Phospho-rac-(1-glycerol)] (Sodium Salt) (DPPG) in layer-by-layer (LbL) films, alternated with poly (amidoamine) G4 (PAMAM) dendrimer layers. The average size of the liposomes in solution was 120 nm as determined by dynamic light scattering, with their spherical shape being inferred from scanning electron microscopy (SEM) in cast films. LbL films containing up to 20 PAMAM/DPPG bilayers were assembled onto glass and/or silicon wafer substrates. The growth of the multilayers was achieved by alternately immersing the substrates into the PAMAM and DPPG solutions for 5 and 10 min, respectively. The formation of PAMAM/DPPG liposome multilayers and its ability to interact with BSA were confirmed by Fourier transform infrared spectroscopy (FTIR). The structural features and film thickness were obtained using X-ray diffraction and surface plasmon resonance (SPR). (c) 2007 Elsevier B.V. All rights reserved.

Identification of lift-off mechanism failure for salt drill-in drilling fluid containing polymer filter cake through adsorption/desorption studies

Drilling fluid`s contact with the productive zone of horizontal or complex wells can reduce well productivity by fluid invasion in the borehole wall. Salted drilling drill-in fluid containing polymers has often been applied in horizontal or complex petroleum wells in the poorly consolidated sandstone reservoirs of the Campos basin, Rio de Janeiro, Brazil. This fluid usually consists of natural polymers such as starch and xanthan gum, which are deposited as a filter cake on the wellbore wall during the drilling. Therefore, the identification of a lift-off mechanism failure, which can be detachment or blistering and pinholing, will enable formulation improvements. increasing the chances of success during filter cake removal in open hole operations. Likewise, knowledge of drill-in drilling fluid adsorption/desorption onto sand can help understand the filter cake-rock adhesion mechanism and consequently filter cake lift-off mechanism failures. The present study aimed to identify the lift-off failure mechanism for this type of fluid filter cake studying adsorption/desorption onto SiO(2) using solutions of natural polymers, lubricants, besides the fluid itself. Ellipsometry was employed to measure this process. The adsorption/desorption studies showed that the adsorbed layer of drilling fluid onto the walls of the rock pores is made up of clusters of polymers, linked by hydrogen bonds, which results in a force of lower cohesion compared to the electrostatic interaction between silica and polymers. Consequently, it was found that the most probable filter cake failure mechanism is rupture (blistering and pinholing), which results in the formation of ducts within the filter cake. (C) 2009 Elsevier B.V. All rights reserved.

Electrostatic layer-by-layer deposition and electrochemical characterization of thin films composed of MnO(2) nanoparticles in a room-temperature ionic liquid

Thin films of MnO(2) nanoparticles were grown using the layer-by-layer method with poly (diallyldimetylammonium) as the intercalated layer. The film growth was followed by UV-vis, electrochemical quartz crystal microbalance (EQCM), and atomic force microscopy. Linear growth due to electrostatic immobilization of layers was observed up to 30 bilayers, but electrical connectivity was maintained only for 12 MnO(2)/PPDA bilayers. The electrochemical characterization of this film in 1-butyl-2,3-dimethyl-imidazolium (BMMI) bis(trifluoromethanesulfonyl)imide (TFSI) (BMMITFSI) with and without addition of a lithium salt indicated a higher electrochemical response of the nanostructured electrode in the lithium-containing electrolyte. On the basis of EQCM experiments, it was possible to confirm that the charge compensation process is achieved mainly by the TFSI anion at short times (<2 s) and by BMMI and lithium cations at longer times. The fact that large ions like TFSI and BMMI participate in the electroneutrality is attributed to the redox reaction that occurs at the superficial sites and to the high concentration of these species compared to that of lithium cations.

Physical quality of an Oxisol cultivated with maize submited to cover crops in the pre-cropping period

The intensive use of land alters the distribution of the pore size which imparts consequences on the soil physical quality. The Least Limiting Water Range (LLWR) allows for the visualization of the effects of management systems upon either the improvement or the degradation of the soil physical quality. The objective of this study was to evaluate the physical quality of a Red Latosol (Oxisol) submited to cover crops in the period prior to the maize crop in a no-tillage and conventional tillage system, using porosity, soil bulk density and the LLWR as attributes. The treatments were: conventional tillage (CT) and a no-tillage system with the following cover crops: sunn hemp (Crotalaria juncea L.) (NS), pearl millet (Pennisetum americanum (L.) Leeke) (NP) and lablab (Dolichos lablab L.) (NL). The experimental design was randomized blocks in subdivided plots with six replications, with the plots being constituted by the treatments and the subplots by the layers analyzed. The no-tillage systems showed higher total porosity and soil organic matter at the 0-0.5 m layer for the CT. The CT did not differ from the NL or NS in relation to macroporosity. The NP showed the greater porosity, while CT and NS presented lower soil bulk density. No <= 10 % airing porosity was found for the treatments evaluated, and value for water content where soil aeration is critical (theta(PA)) was found above estimated water content at field capacity (theta(FC)) for all densities. Critical soil bulk density was of 1.36 and 1.43 Mg m(-3) for NP and CT, respectively. The LLWR in the no-tillage systems was limited in the upper part by the theta(FC), and in the bottom part, by the water content from which soil resistance to penetration is limiting (theta(PR)). By means of LLWR it was observed that the soil presented good physical quality.

Hybrid layer thickness and resin tag length of a self-etching adhesive bonded to sound dentin

Objectives: the purpose of this study is to employ optical microscopy to measure the thickness of the hybrid layer and the penetration (tags) of an aggressive self-etching adhesive system into sound dentin.Methods: occtusat cavities were prepared in 40 extracted human posterior teeth. The prepared teeth were randomly assigned to four experimental groups with 10 specimens each. The self-etching adhesive system Adper Prompt L-Pop was applied to the dentin surface as follows: Group 1: cavosurface enamel was etched for 60 s and dentin for 20 s with 35% phosphoric acid get, immediately followed by application of the self -etching adhesive with a brush to the entire cavity for 15 s; Groups 2, 3, and 4: no pre-etching was performed, and the self -etching adhesive was applied to both enamel and dentin for 15, 30 and 45 s, respectively. After curing, the cavities were fitted with composite resin Fittek Z250. Afterwards, the teeth were decalcified and the restorations were carefully removed for later embedding in paraffin. The specimens were serially sectioned at 6 mu m of thickness and sequentially mounted in glass slides. These sections were stained with Brown and Brenn staining for posterior analysis and measurement of the hybrid layer and resin tags on a tight microscope with a micrometric ocular 40/075. The results were submitted to analysis of variance at the 5% level.Results: whenever there was significance, the Tukey test was applied at the 5% level. The specimens receiving application of acid etching before the selfetching. adhesive displayed a larger thickness of the hybrid layer; on the other hand, specimens receiving only application of the self -etching adhesive on dentin for 15, 30 and 45 s exhibited similar thickness of the hybrid layer. As regards the resin tags, no statistically significant differences could be found between the study groups.Conclusions: it could be concluded that the increase in the time of application of the self-etching adhesive Adper Prompt L-Pop did not significantly influence the formation and thickness of hybrid layer, as well as its penetration into the sound dentin surface. (c) 2005 Elsevier Ltd. All rights reserved.

Dominance and Subordination Interactions Among Nestmates in Pre and Post-Emergence Phases of the Basal Eusocial Wasp Mischocyttarus (Monogynoecus) montei (Hymenoptera, Vespidae)

Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES)