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.

Design and Development of an Adaptable Frame -Based System for Dravidian Language Processing

This work is aimed at building an adaptable frame-based system for processing Dravidian languages. There are about 17 languages in this family and they are spoken by the people of South India.Karaka relations are one of the most important features of Indian languages. They are the semabtuco-syntactic relations between verbs and other related constituents in a sentence. The karaka relations and surface case endings are analyzed for meaning extraction. This approach is comparable with the borad class of case based grammars.The efficiency of this approach is put into test in two applications. One is machine translation and the other is a natural language interface (NLI) for information retrieval from databases. The system mainly consists of a morphological analyzer, local word grouper, a parser for the source language and a sentence generator for the target language. This work make contributios like, it gives an elegant account of the relation between vibhakthi and karaka roles in Dravidian languages. This mapping is elegant and compact. The same basic thing also explains simple and complex sentence in these languages. This suggests that the solution is not just ad hoc but has a deeper underlying unity. This methodology could be extended to other free word order languages. Since the frame designed for meaning representation is general, they are adaptable to other languages coming in this group and to other applications.

Statistical Machine Learning Techniques for the Prediction of Learning Disabilities in School-Age Children

Learning Disability (LD) is a general term that describes specific kinds of learning problems. It is a neurological condition that affects a child's brain and impairs his ability to carry out one or many specific tasks. The learning disabled children are neither slow nor mentally retarded. This disorder can make it problematic for a child to learn as quickly or in the same way as some child who isn't affected by a learning disability. An affected child can have normal or above average intelligence. They may have difficulty paying attention, with reading or letter recognition, or with mathematics. It does not mean that children who have learning disabilities are less intelligent. In fact, many children who have learning disabilities are more intelligent than an average child. Learning disabilities vary from child to child. One child with LD may not have the same kind of learning problems as another child with LD. There is no cure for learning disabilities and they are life-long. However, children with LD can be high achievers and can be taught ways to get around the learning disability. In this research work, data mining using machine learning techniques are used to analyze the symptoms of LD, establish interrelationships between them and evaluate the relative importance of these symptoms. To increase the diagnostic accuracy of learning disability prediction, a knowledge based tool based on statistical machine learning or data mining techniques, with high accuracy,according to the knowledge obtained from the clinical information, is proposed. The basic idea of the developed knowledge based tool is to increase the accuracy of the learning disability assessment and reduce the time used for the same. Different statistical machine learning techniques in data mining are used in the study. Identifying the important parameters of LD prediction using the data mining techniques, identifying the hidden relationship between the symptoms of LD and estimating the relative significance of each symptoms of LD are also the parts of the objectives of this research work. The developed tool has many advantages compared to the traditional methods of using check lists in determination of learning disabilities. For improving the performance of various classifiers, we developed some preprocessing methods for the LD prediction system. A new system based on fuzzy and rough set models are also developed for LD prediction. Here also the importance of pre-processing is studied. A Graphical User Interface (GUI) is designed for developing an integrated knowledge based tool for prediction of LD as well as its degree. The designed tool stores the details of the children in the student database and retrieves their LD report as and when required. The present study undoubtedly proves the effectiveness of the tool developed based on various machine learning techniques. It also identifies the important parameters of LD and accurately predicts the learning disability in school age children. This thesis makes several major contributions in technical, general and social areas. The results are found very beneficial to the parents, teachers and the institutions. They are able to diagnose the child’s problem at an early stage and can go for the proper treatments/counseling at the correct time so as to avoid the academic and social losses.

Dialogue Systems for Language Learning

En este artículo se presenta el caso de Milao, un entorno virtual que ofrece a los estudiantes de idiomas extranjeros la oportunidad de desarrollar y mejorar sus habilidades comunicativas dialogando en escenarios de conversación predefinidos que simulan la interacción con un nativo. Esta tecnología propone una solución a uno de los mayores retos en el aprendizaje de lenguas extranjeras: la falta de oportunidades para poner en práctica la gramática y el vocabulario recién adquiridos. Combinando la investigación sobre la lingüística y el aprendizaje de lenguas con los avances tecnológicos en el campo del Procesamiento del Lenguaje Natural (NPL), particularmente sobre sistemas de diálogo, hemos creado oportunidades en la demanda de los estudiantes a conversar en la lengua que tratan de aprender.

Complex networks analysis of manual and machine translations

Complex networks have been increasingly used in text analysis, including in connection with natural language processing tools, as important text features appear to be captured by the topology and dynamics of the networks. Following previous works that apply complex networks concepts to text quality measurement, summary evaluation, and author characterization, we now focus on machine translation (MT). In this paper we assess the possible representation of texts as complex networks to evaluate cross-linguistic issues inherent in manual and machine translation. We show that different quality translations generated by NIT tools can be distinguished from their manual counterparts by means of metrics such as in-(ID) and out-degrees (OD), clustering coefficient (CC), and shortest paths (SP). For instance, we demonstrate that the average OD in networks of automatic translations consistently exceeds the values obtained for manual ones, and that the CC values of source texts are not preserved for manual translations, but are for good automatic translations. This probably reflects the text rearrangements humans perform during manual translation. We envisage that such findings could lead to better NIT tools and automatic evaluation metrics.

Simultaneous prediction of symptom severity and cause in data from a test battery for Parkinson patients, using machine learning methods

The main purpose of this thesis project is to prediction of symptom severity and cause in data from test battery of the Parkinson’s disease patient, which is based on data mining. The collection of the data is from test battery on a hand in computer. We use the Chi-Square method and check which variables are important and which are not important. Then we apply different data mining techniques on our normalize data and check which technique or method gives good results.The implementation of this thesis is in WEKA. We normalize our data and then apply different methods on this data. The methods which we used are Naïve Bayes, CART and KNN. We draw the Bland Altman and Spearman’s Correlation for checking the final results and prediction of data. The Bland Altman tells how the percentage of our confident level in this data is correct and Spearman’s Correlation tells us our relationship is strong. On the basis of results and analysis we see all three methods give nearly same results. But if we see our CART (J48 Decision Tree) it gives good result of under predicted and over predicted values that’s lies between -2 to +2. The correlation between the Actual and Predicted values is 0,794in CART. Cause gives the better percentage classification result then disability because it can use two classes.

Machine Learning Algorithms in Heavy Process Manufacturing

In a global economy, manufacturers mainly compete with cost efficiency of production, as the price of raw materials are similar worldwide. Heavy industry has two big issues to deal with. On the one hand there is lots of data which needs to be analyzed in an effective manner, and on the other hand making big improvements via investments in cooperate structure or new machinery is neither economically nor physically viable. Machine learning offers a promising way for manufacturers to address both these problems as they are in an excellent position to employ learning techniques with their massive resource of historical production data. However, choosing modelling a strategy in this setting is far from trivial and this is the objective of this article. The article investigates characteristics of the most popular classifiers used in industry today. Support Vector Machines, Multilayer Perceptron, Decision Trees, Random Forests, and the meta-algorithms Bagging and Boosting are mainly investigated in this work. Lessons from real-world implementations of these learners are also provided together with future directions when different learners are expected to perform well. The importance of feature selection and relevant selection methods in an industrial setting are further investigated. Performance metrics have also been discussed for the sake of completion.

Abordagem para o desenvolvimento de um etiquetador de alta acurácia para o Português do Brasil

A etiquetagem morfossintática é uma tarefa básica requerida por muitas aplicações de processamento de linguagem natural, tais como análise gramatical e tradução automática, e por aplicações de processamento de fala, por exemplo, síntese de fala. Essa tarefa consiste em etiquetar palavras em uma sentença com as suas categorias gramaticais. Apesar dessas aplicações requererem etiquetadores que demandem maior precisão, os etiquetadores do estado da arte ainda alcançam acurácia de 96 a 97%. Nesta tese, são investigados recursos de corpus e de software para o desenvolvimento de um etiquetador com acurácia superior à do estado da arte para o português brasileiro. Centrada em uma solução híbrida que combina etiquetagem probabilística com etiquetagem baseada em regras, a proposta de tese se concentra em um estudo exploratório sobre o método de etiquetagem, o tamanho, a qualidade, o conjunto de etiquetas e o gênero dos corpora de treinamento e teste, além de avaliar a desambiguização de palavras novas ou desconhecidas presentes nos textos a serem etiquetados. Quatro corpora foram usados nos experimentos: CETENFolha, Bosque CF 7.4, Mac-Morpho e Selva Científica. O modelo de etiquetagem proposto partiu do uso do método de aprendizado baseado em transformação(TBL) ao qual foram adicionadas três estratégias, combinadas em uma arquitetura que integra as saídas (textos etiquetados) de duas ferramentas de uso livre, o TreeTagger e o -TBL, com os módulos adicionados ao modelo. No modelo de etiquetador treinado com o corpus Mac-Morpho, de gênero jornalístico, foram obtidas taxas de acurácia de 98,05% na etiquetagem de textos do Mac-Morpho e 98,27% em textos do Bosque CF 7.4, ambos de gênero jornalístico. Avaliou-se também o desempenho do modelo de etiquetador híbrido proposto na etiquetagem de textos do corpus Selva Científica, de gênero científico. Foram identificadas necessidades de ajustes no etiquetador e nos corpora e, como resultado, foram alcançadas taxas de acurácia de 98,07% no Selva Científica, 98,06% no conjunto de teste do Mac-Morpho e 98,30% em textos do Bosque CF 7.4. Esses resultados são significativos, pois as taxas de acurácia alcançadas são superiores às do estado da arte, validando o modelo proposto em busca de um etiquetador morfossintático mais confiável.

Using Machine Learning Classifiers to Assist Healthcare-Related Decisions: Classification of Electronic Patient Records

Surveillance Levels (SLs) are categories for medical patients (used in Brazil) that represent different types of medical recommendations. SLs are defined according to risk factors and the medical and developmental history of patients. Each SL is associated with specific educational and clinical measures. The objective of the present paper was to verify computer-aided, automatic assignment of SLs. The present paper proposes a computer-aided approach for automatic recommendation of SLs. The approach is based on the classification of information from patient electronic records. For this purpose, a software architecture composed of three layers was developed. The architecture is formed by a classification layer that includes a linguistic module and machine learning classification modules. The classification layer allows for the use of different classification methods, including the use of preprocessed, normalized language data drawn from the linguistic module. We report the verification and validation of the software architecture in a Brazilian pediatric healthcare institution. The results indicate that selection of attributes can have a great effect on the performance of the system. Nonetheless, our automatic recommendation of surveillance level can still benefit from improvements in processing procedures when the linguistic module is applied prior to classification. Results from our efforts can be applied to different types of medical systems. The results of systems supported by the framework presented in this paper may be used by healthcare and governmental institutions to improve healthcare services in terms of establishing preventive measures and alerting authorities about the possibility of an epidemic.

Deep learning for computer vision: a comparison between convolutional neural networks and hierarchical temporal memories on object recognition tasks

In recent years, Deep Learning techniques have shown to perform well on a large variety of problems both in Computer Vision and Natural Language Processing, reaching and often surpassing the state of the art on many tasks. The rise of deep learning is also revolutionizing the entire field of Machine Learning and Pattern Recognition pushing forward the concepts of automatic feature extraction and unsupervised learning in general. However, despite the strong success both in science and business, deep learning has its own limitations. It is often questioned if such techniques are only some kind of brute-force statistical approaches and if they can only work in the context of High Performance Computing with tons of data. Another important question is whether they are really biologically inspired, as claimed in certain cases, and if they can scale well in terms of "intelligence". The dissertation is focused on trying to answer these key questions in the context of Computer Vision and, in particular, Object Recognition, a task that has been heavily revolutionized by recent advances in the field. Practically speaking, these answers are based on an exhaustive comparison between two, very different, deep learning techniques on the aforementioned task: Convolutional Neural Network (CNN) and Hierarchical Temporal memory (HTM). They stand for two different approaches and points of view within the big hat of deep learning and are the best choices to understand and point out strengths and weaknesses of each of them. CNN is considered one of the most classic and powerful supervised methods used today in machine learning and pattern recognition, especially in object recognition. CNNs are well received and accepted by the scientific community and are already deployed in large corporation like Google and Facebook for solving face recognition and image auto-tagging problems. HTM, on the other hand, is known as a new emerging paradigm and a new meanly-unsupervised method, that is more biologically inspired. It tries to gain more insights from the computational neuroscience community in order to incorporate concepts like time, context and attention during the learning process which are typical of the human brain. In the end, the thesis is supposed to prove that in certain cases, with a lower quantity of data, HTM can outperform CNN.

Shallow Dialogue Processing Using Machine Learning Algorithms (or not)

This paper presents a shallow dialogue analysis model, aimed at human-human dialogues in the context of staff or business meetings. Four components of the model are defined, and several machine learning techniques are used to extract features from dialogue transcripts: maximum entropy classifiers for dialogue acts, latent semantic analysis for topic segmentation, or decision tree classifiers for discourse markers. A rule-based approach is proposed for solving cross-modal references to meeting documents. The methods are trained and evaluated thanks to a common data set and annotation format. The integration of the components into an automated shallow dialogue parser opens the way to multimodal meeting processing and retrieval applications.

BIOMEDICAL LANGUAGE UNDERSTANDING AND EXTRACTION (BLUE-TEXT): A MINIMAL SYNTACTIC, SEMANTIC METHOD

Clinical text understanding (CTU) is of interest to health informatics because critical clinical information frequently represented as unconstrained text in electronic health records are extensively used by human experts to guide clinical practice, decision making, and to document delivery of care, but are largely unusable by information systems for queries and computations. Recent initiatives advocating for translational research call for generation of technologies that can integrate structured clinical data with unstructured data, provide a unified interface to all data, and contextualize clinical information for reuse in multidisciplinary and collaborative environment envisioned by CTSA program. This implies that technologies for the processing and interpretation of clinical text should be evaluated not only in terms of their validity and reliability in their intended environment, but also in light of their interoperability, and ability to support information integration and contextualization in a distributed and dynamic environment. This vision adds a new layer of information representation requirements that needs to be accounted for when conceptualizing implementation or acquisition of clinical text processing tools and technologies for multidisciplinary research. On the other hand, electronic health records frequently contain unconstrained clinical text with high variability in use of terms and documentation practices, and without commitmentto grammatical or syntactic structure of the language (e.g. Triage notes, physician and nurse notes, chief complaints, etc). This hinders performance of natural language processing technologies which typically rely heavily on the syntax of language and grammatical structure of the text. This document introduces our method to transform unconstrained clinical text found in electronic health information systems to a formal (computationally understandable) representation that is suitable for querying, integration, contextualization and reuse, and is resilient to the grammatical and syntactic irregularities of the clinical text. We present our design rationale, method, and results of evaluation in processing chief complaints and triage notes from 8 different emergency departments in Houston Texas. At the end, we will discuss significance of our contribution in enabling use of clinical text in a practical bio-surveillance setting.

Método para la Construcción Automática de Ontologías Basado en Patrones Lingüísticos

La tesis que se presenta tiene como propósito la construcción automática de ontologías a partir de textos, enmarcándose en el área denominada Ontology Learning. Esta disciplina tiene como objetivo automatizar la elaboración de modelos de dominio a partir de fuentes información estructurada o no estructurada, y tuvo su origen con el comienzo del milenio, a raíz del crecimiento exponencial del volumen de información accesible en Internet. Debido a que la mayoría de información se presenta en la web en forma de texto, el aprendizaje automático de ontologías se ha centrado en el análisis de este tipo de fuente, nutriéndose a lo largo de los años de técnicas muy diversas provenientes de áreas como la Recuperación de Información, Extracción de Información, Sumarización y, en general, de áreas relacionadas con el procesamiento del lenguaje natural. La principal contribución de esta tesis consiste en que, a diferencia de la mayoría de las técnicas actuales, el método que se propone no analiza la estructura sintáctica superficial del lenguaje, sino que estudia su nivel semántico profundo. Su objetivo, por tanto, es tratar de deducir el modelo del dominio a partir de la forma con la que se articulan los significados de las oraciones en lenguaje natural. Debido a que el nivel semántico profundo es independiente de la lengua, el método permitirá operar en escenarios multilingües, en los que es necesario combinar información proveniente de textos en diferentes idiomas. Para acceder a este nivel del lenguaje, el método utiliza el modelo de las interlinguas. Estos formalismos, provenientes del área de la traducción automática, permiten representar el significado de las oraciones de forma independiente de la lengua. Se utilizará en concreto UNL (Universal Networking Language), considerado como la única interlingua de propósito general que está normalizada. La aproximación utilizada en esta tesis supone la continuación de trabajos previos realizados tanto por su autor como por el equipo de investigación del que forma parte, en los que se estudió cómo utilizar el modelo de las interlinguas en las áreas de extracción y recuperación de información multilingüe. Básicamente, el procedimiento definido en el método trata de identificar, en la representación UNL de los textos, ciertas regularidades que permiten deducir las piezas de la ontología del dominio. Debido a que UNL es un formalismo basado en redes semánticas, estas regularidades se presentan en forma de grafos, generalizándose en estructuras denominadas patrones lingüísticos. Por otra parte, UNL aún conserva ciertos mecanismos de cohesión del discurso procedentes de los lenguajes naturales, como el fenómeno de la anáfora. Con el fin de aumentar la efectividad en la comprensión de las expresiones, el método provee, como otra contribución relevante, la definición de un algoritmo para la resolución de la anáfora pronominal circunscrita al modelo de la interlingua, limitada al caso de pronombres personales de tercera persona cuando su antecedente es un nombre propio. El método propuesto se sustenta en la definición de un marco formal, que ha debido elaborarse adaptando ciertas definiciones provenientes de la teoría de grafos e incorporando otras nuevas, con el objetivo de ubicar las nociones de expresión UNL, patrón lingüístico y las operaciones de encaje de patrones, que son la base de los procesos del método. Tanto el marco formal como todos los procesos que define el método se han implementado con el fin de realizar la experimentación, aplicándose sobre un artículo de la colección EOLSS “Encyclopedia of Life Support Systems” de la UNESCO. ABSTRACT The purpose of this thesis is the automatic construction of ontologies from texts. This thesis is set within the area of Ontology Learning. This discipline aims to automatize domain models from structured or unstructured information sources, and had its origin with the beginning of the millennium, as a result of the exponential growth in the volume of information accessible on the Internet. Since most information is presented on the web in the form of text, the automatic ontology learning is focused on the analysis of this type of source, nourished over the years by very different techniques from areas such as Information Retrieval, Information Extraction, Summarization and, in general, by areas related to natural language processing. The main contribution of this thesis consists of, in contrast with the majority of current techniques, the fact that the method proposed does not analyze the syntactic surface structure of the language, but explores his deep semantic level. Its objective, therefore, is trying to infer the domain model from the way the meanings of the sentences are articulated in natural language. Since the deep semantic level does not depend on the language, the method will allow to operate in multilingual scenarios, where it is necessary to combine information from texts in different languages. To access to this level of the language, the method uses the interlingua model. These formalisms, coming from the area of machine translation, allow to represent the meaning of the sentences independently of the language. In this particular case, UNL (Universal Networking Language) will be used, which considered to be the only interlingua of general purpose that is standardized. The approach used in this thesis corresponds to the continuation of previous works carried out both by the author of this thesis and by the research group of which he is part, in which it is studied how to use the interlingua model in the areas of multilingual information extraction and retrieval. Basically, the procedure defined in the method tries to identify certain regularities at the UNL representation of texts that allow the deduction of the parts of the ontology of the domain. Since UNL is a formalism based on semantic networks, these regularities are presented in the form of graphs, generalizing in structures called linguistic patterns. On the other hand, UNL still preserves certain mechanisms of discourse cohesion from natural languages, such as the phenomenon of the anaphora. In order to increase the effectiveness in the understanding of expressions, the method provides, as another significant contribution, the definition of an algorithm for the resolution of pronominal anaphora limited to the model of the interlingua, in the case of third person personal pronouns when its antecedent is a proper noun. The proposed method is based on the definition of a formal framework, adapting some definitions from Graph Theory and incorporating new ones, in order to locate the notions of UNL expression and linguistic pattern, as well as the operations of pattern matching, which are the basis of the method processes. Both the formal framework and all the processes that define the method have been implemented in order to carry out the experimentation, applying on an article of the "Encyclopedia of Life Support Systems" of the UNESCO-EOLSS collection.

Human Immunodeficiency Virus Reverse Transcriptase and Protease Sequence Database: an expanded data model integrating natural language text and sequence analysis programs

The HIV Reverse Transcriptase and Protease Sequence Database is an on-line relational database that catalogs evolutionary and drug-related sequence variation in the human immunodeficiency virus (HIV) reverse transcriptase (RT) and protease enzymes, the molecular targets of anti-HIV therapy (http://hivdb.stanford.edu). The database contains a compilation of nearly all published HIV RT and protease sequences, including submissions from International Collaboration databases and sequences published in journal articles. Sequences are linked to data about the source of the sequence sample and the antiretroviral drug treatment history of the individual from whom the isolate was obtained. During the past year 3500 sequences have been added and the data model has been expanded to include drug susceptibility data on sequenced isolates. Database content has also been integrated with didactic text and the output of two sequence analysis programs.

Structure-activity relationships derived by machine learning: the use of atoms and their bond connectivities to predict mutagenicity by inductive logic programming.

We present a general approach to forming structure-activity relationships (SARs). This approach is based on representing chemical structure by atoms and their bond connectivities in combination with the inductive logic programming (ILP) algorithm PROGOL. Existing SAR methods describe chemical structure by using attributes which are general properties of an object. It is not possible to map chemical structure directly to attribute-based descriptions, as such descriptions have no internal organization. A more natural and general way to describe chemical structure is to use a relational description, where the internal construction of the description maps that of the object described. Our atom and bond connectivities representation is a relational description. ILP algorithms can form SARs with relational descriptions. We have tested the relational approach by investigating the SARs of 230 aromatic and heteroaromatic nitro compounds. These compounds had been split previously into two subsets, 188 compounds that were amenable to regression and 42 that were not. For the 188 compounds, a SAR was found that was as accurate as the best statistical or neural network-generated SARs. The PROGOL SAR has the advantages that it did not need the use of any indicator variables handcrafted by an expert, and the generated rules were easily comprehensible. For the 42 compounds, PROGOL formed a SAR that was significantly (P < 0.025) more accurate than linear regression, quadratic regression, and back-propagation. This SAR is based on an automatically generated structural alert for mutagenicity.