Multi-Model Data Fusion to Improve an Early Warning System for Hypo-/Hyperglycemic Events

Correct predictions of future blood glucose levels in individuals with Type 1 Diabetes (T1D) can be used to provide early warning of upcoming hypo-/hyperglycemic events and thus to improve the patient's safety. To increase prediction accuracy and efficiency, various approaches have been proposed which combine multiple predictors to produce superior results compared to single predictors. Three methods for model fusion are presented and comparatively assessed. Data from 23 T1D subjects under sensor-augmented pump (SAP) therapy were used in two adaptive data-driven models (an autoregressive model with output correction - cARX, and a recurrent neural network - RNN). Data fusion techniques based on i) Dempster-Shafer Evidential Theory (DST), ii) Genetic Algorithms (GA), and iii) Genetic Programming (GP) were used to merge the complimentary performances of the prediction models. The fused output is used in a warning algorithm to issue alarms of upcoming hypo-/hyperglycemic events. The fusion schemes showed improved performance with lower root mean square errors, lower time lags, and higher correlation. In the warning algorithm, median daily false alarms (DFA) of 0.25%, and 100% correct alarms (CA) were obtained for both event types. The detection times (DT) before occurrence of events were 13.0 and 12.1 min respectively for hypo-/hyperglycemic events. Compared to the cARX and RNN models, and a linear fusion of the two, the proposed fusion schemes represents a significant improvement.

Contexto e modularização em redes neurais recorrentes para aprendizagem de seqüências temporais

Este trabalho apresenta um sistema neural modular, que processa separadamente informações de contexto espacial e temporal, para a tarefa de reprodução de sequências temporais. Para o desenvolvimento do sistema neural foram considerados redes neurais recorrentes, modelos estocásticos, sistemas neurais modulares e processamento de informações de contexto. Em seguida, foram estudados três modelos com abordagens distintas para aprendizagem de seqüências temporais: uma rede neural parcialmente recorrente, um exemplo de sistema neural modular e um modelo estocástico utilizando a teoria de modelos markovianos escondidos. Com base nos estudos e modelos apresentados, esta pesquisa propõe um sistema formado por dois módulos sucessivos distintos. Uma rede de propagação direta (módulo estimador de contexto espacial) realiza o processamento de contexto espacial identificando a seqüência a ser reproduzida e fornecendo um protótipo do contexto para o segundo módulo. Este é formado por uma rede parcialmente recorrente (módulo de reprodução de sequências temporais) para aprender as informações de contexto temporal e reproduzir em suas saídas a seqüência identificada pelo módulo anterior. Para a finalidade mencionada, este mestrado utiliza a distribuição de Gibbs na saída do módulo para contexto espacial de forma que este forneça probabilidades de contexto espacial, indicando o grau de certeza do módulo e possibilitando a utilização de procedimentos especiais para os casos de dúvida. O sistema neural foi testado em conjuntos contendo trajetórias abertas, fechadas, e com diferentes situações de ambigüidade e complexidade. Duas situações distintas foram avaliadas: (a) capacidade do sistema em reproduzir trajetórias a partir de pontos iniciais treinados; e (b) capacidade de generalização do sistema reproduzindo trajetórias considerando pontos iniciais ou finais em situações não treinadas. A situação (b) é um problema de difícil ) solução em redes neurais devido à falta de contexto temporal, essencial na reprodução de seqüências. Foram realizados experimentos comparando o desempenho do sistema modular proposto com o de uma rede parcialmente recorrente operando sozinha e um sistema modular neural (TOTEM). Os resultados sugerem que o sistema proposto apresentou uma capacidade de generalização significamente melhor, sem que houvesse uma deterioração na capacidade de reproduzir seqüências treinadas. Esses resultados foram obtidos em sistema mais simples que o TOTEM.

Learning user and product distributed representations using a sequence model for sentiment analysis

In product reviews, it is observed that the distribution of polarity ratings over reviews written by different users or evaluated based on different products are often skewed in the real world. As such, incorporating user and product information would be helpful for the task of sentiment classification of reviews. However, existing approaches ignored the temporal nature of reviews posted by the same user or evaluated on the same product. We argue that the temporal relations of reviews might be potentially useful for learning user and product embedding and thus propose employing a sequence model to embed these temporal relations into user and product representations so as to improve the performance of document-level sentiment analysis. Specifically, we first learn a distributed representation of each review by a one-dimensional convolutional neural network. Then, taking these representations as pretrained vectors, we use a recurrent neural network with gated recurrent units to learn distributed representations of users and products. Finally, we feed the user, product and review representations into a machine learning classifier for sentiment classification. Our approach has been evaluated on three large-scale review datasets from the IMDB and Yelp. Experimental results show that: (1) sequence modeling for the purposes of distributed user and product representation learning can improve the performance of document-level sentiment classification; (2) the proposed approach achieves state-of-The-Art results on these benchmark datasets.

Estudo e implementação de algoritmos inteligentes para detecção e classificação de falhas na medição de gás natural

This master dissertation presents the study and implementation of inteligent algorithms to monitor the measurement of sensors involved in natural gas custody transfer processes. To create these algoritmhs Artificial Neural Networks are investigated because they have some particular properties, such as: learning, adaptation, prediction. A neural predictor is developed to reproduce the sensor output dynamic behavior, in such a way that its output is compared to the real sensor output. A recurrent neural network is used for this purpose, because of its ability to deal with dynamic information. The real sensor output and the estimated predictor output work as the basis for the creation of possible sensor fault detection and diagnosis strategies. Two competitive neural network architectures are investigated and their capabilities are used to classify different kinds of faults. The prediction algorithm and the fault detection classification strategies, as well as the obtained results, are presented

Development of low entropy coding in a recurrent network.

In this paper we present an unsupervised neural network which exhibits competition between units via inhibitory feedback. The operation is such as to minimize reconstruction error, both for individual patterns, and over the entire training set. A key difference from networks which perform principal components analysis, or one of its variants, is the ability to converge to non-orthogonal weight values. We discuss the network's operation in relation to the twin goals of maximizing information transfer and minimizing code entropy, and show how the assignment of prior probabilities to network outputs can help to reduce entropy. We present results from two binary coding problems, and from experiments with image coding.

Extracted Memory from Temporal Patterns Using Adaptive Resonance and Recurrent Network

We can recognize objects through receiving continuously huge temporal information including redundancy and noise, and can memorize them. This paper proposes a neural network model which extracts pre-recognized patterns from temporally sequential patterns which include redundancy, and memorizes the patterns temporarily. This model consists of an adaptive resonance system and a recurrent time-delay network. The extraction is executed by the matching mechanism of the adaptive resonance system, and the temporal information is processed and stored by the recurrent network. Simple simulations are examined to exemplify the property of extraction.

The sample complexity of pattern classification with neural networks: the size of the weights is more important than the size of the network

Sample complexity results from computational learning theory, when applied to neural network learning for pattern classification problems, suggest that for good generalization performance the number of training examples should grow at least linearly with the number of adjustable parameters in the network. Results in this paper show that if a large neural network is used for a pattern classification problem and the learning algorithm finds a network with small weights that has small squared error on the training patterns, then the generalization performance depends on the size of the weights rather than the number of weights. For example, consider a two-layer feedforward network of sigmoid units, in which the sum of the magnitudes of the weights associated with each unit is bounded by A and the input dimension is n. We show that the misclassification probability is no more than a certain error estimate (that is related to squared error on the training set) plus A3 √((log n)/m) (ignoring log A and log m factors), where m is the number of training patterns. This may explain the generalization performance of neural networks, particularly when the number of training examples is considerably smaller than the number of weights. It also supports heuristics (such as weight decay and early stopping) that attempt to keep the weights small during training. The proof techniques appear to be useful for the analysis of other pattern classifiers: when the input domain is a totally bounded metric space, we use the same approach to give upper bounds on misclassification probability for classifiers with decision boundaries that are far from the training examples.

Using strategic movement to calibrate a neural compass : a spiking network for tracking head direction in rats and robots

The head direction (HD) system in mammals contains neurons that fire to represent the direction the animal is facing in its environment. The ability of these cells to reliably track head direction even after the removal of external sensory cues implies that the HD system is calibrated to function effectively using just internal (proprioceptive and vestibular) inputs. Rat pups and other infant mammals display stereotypical warm-up movements prior to locomotion in novel environments, and similar warm-up movements are seen in adult mammals with certain brain lesion-induced motor impairments. In this study we propose that synaptic learning mechanisms, in conjunction with appropriate movement strategies based on warm-up movements, can calibrate the HD system so that it functions effectively even in darkness. To examine the link between physical embodiment and neural control, and to determine that the system is robust to real-world phenomena, we implemented the synaptic mechanisms in a spiking neural network and tested it on a mobile robot platform. Results show that the combination of the synaptic learning mechanisms and warm-up movements are able to reliably calibrate the HD system so that it accurately tracks real-world head direction, and that calibration breaks down in systematic ways if certain movements are omitted. This work confirms that targeted, embodied behaviour can be used to calibrate neural systems, demonstrates that ‘grounding’ of modeled biological processes in the real world can reveal underlying functional principles (supporting the importance of robotics to biology), and proposes a functional role for stereotypical behaviours seen in infant mammals and those animals with certain motor deficits. We conjecture that these calibration principles may extend to the calibration of other neural systems involved in motion tracking and the representation of space, such as grid cells in entorhinal cortex.

Genetically Generated Neural Networks I: Representational Effects

This paper studies several applications of genetic algorithms (GAs) within the neural networks field. After generating a robust GA engine, the system was used to generate neural network circuit architectures. This was accomplished by using the GA to determine the weights in a fully interconnected network. The importance of the internal genetic representation was shown by testing different approaches. The effects in speed of optimization of varying the constraints imposed upon the desired network were also studied. It was observed that relatively loose constraints provided results comparable to a fully constrained system. The type of neural network circuits generated were recurrent competitive fields as described by Grossberg (1982).

On Recurrent and Deep Neural Networks

L'apprentissage profond est un domaine de recherche en forte croissance en apprentissage automatique qui est parvenu à des résultats impressionnants dans différentes tâches allant de la classification d'images à la parole, en passant par la modélisation du langage. Les réseaux de neurones récurrents, une sous-classe d'architecture profonde, s'avèrent particulièrement prometteurs. Les réseaux récurrents peuvent capter la structure temporelle dans les données. Ils ont potentiellement la capacité d'apprendre des corrélations entre des événements éloignés dans le temps et d'emmagasiner indéfiniment des informations dans leur mémoire interne. Dans ce travail, nous tentons d'abord de comprendre pourquoi la profondeur est utile. Similairement à d'autres travaux de la littérature, nos résultats démontrent que les modèles profonds peuvent être plus efficaces pour représenter certaines familles de fonctions comparativement aux modèles peu profonds. Contrairement à ces travaux, nous effectuons notre analyse théorique sur des réseaux profonds acycliques munis de fonctions d'activation linéaires par parties, puisque ce type de modèle est actuellement l'état de l'art dans différentes tâches de classification. La deuxième partie de cette thèse porte sur le processus d'apprentissage. Nous analysons quelques techniques d'optimisation proposées récemment, telles l'optimisation Hessian free, la descente de gradient naturel et la descente des sous-espaces de Krylov. Nous proposons le cadre théorique des méthodes à région de confiance généralisées et nous montrons que plusieurs de ces algorithmes développés récemment peuvent être vus dans cette perspective. Nous argumentons que certains membres de cette famille d'approches peuvent être mieux adaptés que d'autres à l'optimisation non convexe. La dernière partie de ce document se concentre sur les réseaux de neurones récurrents. Nous étudions d'abord le concept de mémoire et tentons de répondre aux questions suivantes: Les réseaux récurrents peuvent-ils démontrer une mémoire sans limite? Ce comportement peut-il être appris? Nous montrons que cela est possible si des indices sont fournis durant l'apprentissage. Ensuite, nous explorons deux problèmes spécifiques à l'entraînement des réseaux récurrents, à savoir la dissipation et l'explosion du gradient. Notre analyse se termine par une solution au problème d'explosion du gradient qui implique de borner la norme du gradient. Nous proposons également un terme de régularisation conçu spécifiquement pour réduire le problème de dissipation du gradient. Sur un ensemble de données synthétique, nous montrons empiriquement que ces mécanismes peuvent permettre aux réseaux récurrents d'apprendre de façon autonome à mémoriser des informations pour une période de temps indéfinie. Finalement, nous explorons la notion de profondeur dans les réseaux de neurones récurrents. Comparativement aux réseaux acycliques, la définition de profondeur dans les réseaux récurrents est souvent ambiguë. Nous proposons différentes façons d'ajouter de la profondeur dans les réseaux récurrents et nous évaluons empiriquement ces propositions.

“Reconstruction of Gene Regulatory Network from Expression Profile of Plasma RNA Data of Colorectal Cancer Patients using Soft Computing Techniques”

Microarray data analysis is one of data mining tool which is used to extract meaningful information hidden in biological data. One of the major focuses on microarray data analysis is the reconstruction of gene regulatory network that may be used to provide a broader understanding on the functioning of complex cellular systems. Since cancer is a genetic disease arising from the abnormal gene function, the identification of cancerous genes and the regulatory pathways they control will provide a better platform for understanding the tumor formation and development. The major focus of this thesis is to understand the regulation of genes responsible for the development of cancer, particularly colorectal cancer by analyzing the microarray expression data. In this thesis, four computational algorithms namely fuzzy logic algorithm, modified genetic algorithm, dynamic neural fuzzy network and Takagi Sugeno Kang-type recurrent neural fuzzy network are used to extract cancer specific gene regulatory network from plasma RNA dataset of colorectal cancer patients. Plasma RNA is highly attractive for cancer analysis since it requires a collection of small amount of blood and it can be obtained at any time in repetitive fashion allowing the analysis of disease progression and treatment response.

Using Recurrent Networks for Dimensionality Reduction

This report explores how recurrent neural networks can be exploited for learning high-dimensional mappings. Since recurrent networks are as powerful as Turing machines, an interesting question is how recurrent networks can be used to simplify the problem of learning from examples. The main problem with learning high-dimensional functions is the curse of dimensionality which roughly states that the number of examples needed to learn a function increases exponentially with input dimension. This thesis proposes a way of avoiding this problem by using a recurrent network to decompose a high-dimensional function into many lower dimensional functions connected in a feedback loop.

Inverse model control using recurrent networks

This paper illustrates how internal model control of nonlinear processes can be achieved by recurrent neural networks, e.g. fully connected Hopfield networks. It is shown that using results developed by Kambhampati et al. (1995), that once a recurrent network model of a nonlinear system has been produced, a controller can be produced which consists of the network comprising the inverse of the model and a filter. Thus, the network providing control for the nonlinear system does not require any training after it has been trained to model the nonlinear system. Stability and other issues of importance for nonlinear control systems are also discussed.

The relative order and inverses of recurrent networks

Differential geometry is used to investigate the structure of neural-network-based control systems. The key aspect is relative order—an invariant property of dynamic systems. Finite relative order allows the specification of a minimal architecture for a recurrent network. Any system with finite relative order has a left inverse. It is shown that a recurrent network with finite relative order has a local inverse that is also a recurrent network with the same weights. The results have implications for the use of recurrent networks in the inverse-model-based control of nonlinear systems.

Relative order defines a topology for recurrent networks

This paper uses techniques from control theory in the analysis of trained recurrent neural networks. Differential geometry is used as a framework, which allows the concept of relative order to be applied to neural networks. Any system possessing finite relative order has a left-inverse. Any recurrent network with finite relative order also has an inverse, which is shown to be a recurrent network.