Добірка наукової літератури з теми "Incremental neural network"

Vector Quantization (VQ) is a classic optimization problem and a simple approach to pattern recognition. Applications include lossy data compression, clustering and speech and speaker recognition. Although VQ has largely been replaced by time-aware techniques like Hidden Markov Models (HMMs) and Dynamic Time Warping (DTW) in some applications, such as speech and speaker recognition, VQ still retains some significance due to its much lower computational cost — especially for embedded systems. A recent study also demonstrates a multi-section VQ system which achieves performance rivaling that of DTW in an application to handwritten signature recognition, at a much lower computational cost. Adding sensitivity to temporal patterns to a VQ algorithm could help improve such results further. SOTPAR2 is such an extension of Neural Gas, an Artificial Neural Network algorithm for VQ. SOTPAR2 uses a conceptually simple approach, based on adding lateral connections between network nodes and creating “temporal activity” that diffuses through adjacent nodes. The activity in turn makes the nearest-neighbor classifier biased toward network nodes with high activity, and the SOTPAR2 authors report improvements over Neural Gas in an application to time series prediction. This report presents an investigation of how this same extension affects quantization and prediction performance of the self-organizing incremental neural network (SOINN) algorithm. SOINN is a VQ algorithm which automatically chooses a suitable codebook size and can also be used for clustering with arbitrary cluster shapes. This extension is found to not improve the performance of SOINN, in fact it makes performance worse in all experiments attempted. A discussion of this result is provided, along with a discussion of the impact of the algorithm parameters, and possible future work to improve the results is suggested.<br>Vektorkvantisering (VQ; eng: Vector Quantization) är ett klassiskt problem och en enkel metod för mönsterigenkänning. Bland tillämpningar finns förstörande datakompression, klustring och igenkänning av tal och talare. Även om VQ i stort har ersatts av tidsmedvetna tekniker såsom dolda Markovmodeller (HMM, eng: Hidden Markov Models) och dynamisk tidskrökning (DTW, eng: Dynamic Time Warping) i vissa tillämpningar, som tal- och talarigenkänning, har VQ ännu viss relevans tack vare sin mycket lägre beräkningsmässiga kostnad — särskilt för exempelvis inbyggda system. En ny studie demonstrerar också ett VQ-system med flera sektioner som åstadkommer prestanda i klass med DTW i en tillämpning på igenkänning av handskrivna signaturer, men till en mycket lägre beräkningsmässig kostnad. Att dra nytta av temporala mönster i en VQ-algoritm skulle kunna hjälpa till att förbättra sådana resultat ytterligare. SOTPAR2 är en sådan utökning av Neural Gas, en artificiell neural nätverk-algorithm för VQ. SOTPAR2 använder en konceptuellt enkel idé, baserad på att lägga till sidleds anslutningar mellan nätverksnoder och skapa “temporal aktivitet” som diffunderar genom anslutna noder. Aktiviteten gör sedan så att närmaste-granne-klassificeraren föredrar noder med hög aktivitet, och författarna till SOTPAR2 rapporterar förbättrade resultat jämfört med Neural Gas i en tillämpning på förutsägning av en tidsserie. I denna rapport undersöks hur samma utökning påverkar kvantiserings- och förutsägningsprestanda hos algoritmen självorganiserande inkrementellt neuralt nätverk (SOINN, eng: self-organizing incremental neural network). SOINN är en VQ-algorithm som automatiskt väljer en lämplig kodboksstorlek och också kan användas för klustring med godtyckliga klusterformer. Experimentella resultat visar att denna utökning inte förbättrar prestandan hos SOINN, istället försämrades prestandan i alla experiment som genomfördes. Detta resultat diskuteras, liksom inverkan av parametervärden på prestandan, och möjligt framtida arbete för att förbättra resultaten föreslås.

Este trabalho apresenta um novo modelo de redes neurais para análise e previsão de séries temporais: o modelo ARMA-CIGMN (do inglês, Autoregressive Moving Average Classical Incremental Gaussian Mixture Network) além dos resultados obtidos pelo mesmo. Este modelo se baseia em modificações realizadas em uma versão reformulada da IGMN. A IGMN Clássica, CIGMN, é similar à versão original da IGMN, porém baseada em uma abordagem estatística clássica, a qual também é apresentada neste trabalho. As modificações do algoritmo da IGMN foram feitas para melhor adpatação a séries temporais. O modelo ARMA-CIGMN demonstra boa capacidade preditiva e a modelagem ainda pode ser auxiliada por conhecidas ferramentas estatísticas como a função de autorrelação (acf, do original em inglês autocorrelation function) e a de autocorrelação parcial (pacf, do original em inglês partial autocorrelation function), já utilizadas em modelagem de séries temporais e nos modelos da IGMN original. As comparações foram feitas utilizando-se séries conhecidas e dados simulados. Foram selecionados para comparação os modelos estatísticos clássicos ARIMA (do inglês, Autoregressive Integrated Moving Average), a IGMN original e duas modificações feitas ainda na IGMN original:(i) um modelo similar ao modelo ARMA (do inglês, Autoregressive Moving Average) clássico e (ii) um modelo similar ao modelo NOE (do inglês, Nonlinear Output Error). Também é apresentada um versão reformulada da IGMN, usando a abordagem clássica da estatística, necessária para o desenvolvimento do modelo ARMA-CIGMN.<br>This work presents a new model of neural network for time series analysis and forecasting: the ARMA-CIGMN (Autoregressive Moving Average Classical Incremental Gaussian Mixture Network) model and its analysis. This model is based on modifications made to a reformulated IGMN, the Classical IGMN (CIGMN). The CIGMN is similar to the original IGMN, but based on a classical statistical approach. The modifications to the IGMN algorithm were made to better fit it to time series. The proposed ARMA-CIGMN model demonstrates good forecasts and the modeling procedure can also be aided by known statistical tools as the autocorrelation (acf) and partial autocorrelation functions (pacf), already used in classical statistical time series modeling and also with the original IGMN algorithm models. The ARMA-CIGMN model was evaluated using known series and simulated data. The models used for comparisons were the classical statistical ARIMA model and its variants, the original IGMN and two modifications over the original IGMN: (i) a modification similar to a classical ARMA (Autoregressive Moving Average) model and (ii) a similar NOE (Nonlinear Output Error) model. It is also presented a reformulated IGMN version with a classical statistical approach, which is needed for the ARMA-CIGMN model.

Nous nous intéressons au problème de l'apprentissage continu de réseaux de neurones artificiels dans le cas où les données ne sont accessibles que pour une seule catégorie à la fois. Pour remédier au problème de l'oubli catastrophique qui limite les performances d'apprentissage dans ces conditions, nous proposons une approche basée sur la représentation des données d'une catégorie par une loi normale. Les transformations associées à ces représentations sont effectuées à l'aide de réseaux inversibles, qui peuvent alors être entraînés avec les données d'une seule catégorie. Chaque catégorie se voit attribuer un réseau pour représenter ses caractéristiques. Prédire la catégorie revient alors à identifier le réseau le plus représentatif. L'avantage d'une telle approche est qu'une fois qu'un réseau est entraîné, il n'est plus nécessaire de le mettre à jour par la suite, chaque réseau étant indépendant des autres. C'est cette propriété particulièrement avantageuse qui démarque notre méthode des précédents travaux dans ce domaine. Nous appuyons notre démonstration sur des expériences réalisées sur divers jeux de données et montrons que notre approche fonctionne favorablement comparé à l'état de l'art. Dans un second temps, nous proposons d'optimiser notre approche en réduisant son impact en mémoire en factorisant les paramètres des réseaux. Il est alors possible de réduire significativement le coût de stockage de ces réseaux avec une perte de performances limitée. Enfin, nous étudions également des stratégies pour produire des réseaux capables d'être réutilisés sur le long terme et nous montrons leur pertinence par rapport aux réseaux traditionnellement utilisés pour l'apprentissage continu<br>We are interested in the problem of continual learning of artificial neural networks in the case where the data are available for only one class at a time. To address the problem of catastrophic forgetting that restrain the learning performances in these conditions, we propose an approach based on the representation of the data of a class by a normal distribution. The transformations associated with these representations are performed using invertible neural networks, which can be trained with the data of a single class. Each class is assigned a network that will model its features. In this setting, predicting the class of a sample corresponds to identifying the network that best fit the sample. The advantage of such an approach is that once a network is trained, it is no longer necessary to update it later, as each network is independent of the others. It is this particularly advantageous property that sets our method apart from previous work in this area. We support our demonstration with experiments performed on various datasets and show that our approach performs favorably compared to the state of the art. Subsequently, we propose to optimize our approach by reducing its impact on memory by factoring the network parameters. It is then possible to significantly reduce the storage cost of these networks with a limited performance loss. Finally, we also study strategies to produce efficient feature extractor models for continual learning and we show their relevance compared to the networks traditionally used for continual learning

Self-adaptive systems have limited time to adjust their configurations whenever their adaptation goals, i.e., quality requirements, are violated due to some runtime uncertainties. Within the available time, they need to analyze their adaptation space, i.e., a set of configurations, to find the best adaptation option, i.e., configuration, that can achieve their adaptation goals. Existing formal analysis approaches find the best adaptation option by analyzing the entire adaptation space. However, exhaustive analysis requires time and resources and is therefore only efficient when the adaptation space is small. The size of the adaptation space is often in hundreds or thousands, which makes formal analysis approaches inefficient in large-scale self-adaptive systems. In this thesis, we tackle this problem by presenting an online learning approach that enables formal analysis approaches to analyze large adaptation spaces efficiently. The approach integrates with the standard feedback loop and reduces the adaptation space to a subset of adaptation options that are relevant to the current runtime uncertainties. The subset is then analyzed by the formal analysis approaches, which allows them to complete the analysis faster and efficiently within the available time. We evaluate our approach on two different instances of an Internet of Things application. The evaluation shows that our approach dramatically reduces the adaptation space and analysis time without compromising the adaptation goals.

In recent years, deep learning techniques received great attention in the field of information technology. These techniques proved to be particularly useful and effective in domains like natural language processing, speech recognition and computer vision. In several real world applications deep learning approaches improved the state-of-the-art. In the field of machine learning, deep learning was a real revolution and a number of effective techniques have been proposed for both supervised and unsupervised learning and for representation learning. This thesis focuses on deep learning for object recognition, and in particular, it addresses incremental learning techniques. With incremental learning we denote approaches able to create an initial model from a small training set and to improve the model as new data are available. Using temporal coherent sequences proved to be useful for incremental learning since temporal coherence also allows to operate in unsupervised manners. A critical point of incremental learning is called forgetting which is the risk to forget previously learned patterns as new data are presented. In the first chapters of this work we introduce the basic theory on neural networks, Convolutional Neural Networks and incremental learning. CNN is today one of the most effective approaches for supervised object recognition; it is well accepted by the scientific community and largely used by ICT big players like Google and Facebook: relevant applications are Facebook face recognition and Google image search. The scientific community has several (large) datasets (e.g., ImageNet) for the development and evaluation of object recognition approaches. However very few temporally coherent datasets are available to study incremental approaches. For this reason we decided to collect a new dataset named TCD4R (Temporal Coherent Dataset For Robotics).

A contribução original desta tese é um novo algoritmo que integra um aproximador de funções com alta eficiência amostral com aprendizagem por reforço em espaços de estados contínuos. A pesquisa completa inclui o desenvolvimento de um algoritmo online e incremental capaz de aprender por meio de uma única passada sobre os dados. Este algoritmo, chamado de Fast Incremental Gaussian Mixture Network (FIGMN) foi empregado como um aproximador de funções eficiente para o espaço de estados de tarefas contínuas de aprendizagem por reforço, que, combinado com Q-learning linear, resulta em performance competitiva. Então, este mesmo aproximador de funções foi empregado para modelar o espaço conjunto de estados e valores Q, todos em uma única FIGMN, resultando em um algoritmo conciso e com alta eficiência amostral, i.e., um algoritmo de aprendizagem por reforço capaz de aprender por meio de pouquíssimas interações com o ambiente. Um único episódio é suficiente para aprender as tarefas investigadas na maioria dos experimentos. Os resultados são analisados a fim de explicar as propriedades do algoritmo obtido, e é observado que o uso da FIGMN como aproximador de funções oferece algumas importantes vantagens para aprendizagem por reforço em relação a redes neurais convencionais.<br>This thesis’ original contribution is a novel algorithm which integrates a data-efficient function approximator with reinforcement learning in continuous state spaces. The complete research includes the development of a scalable online and incremental algorithm capable of learning from a single pass through data. This algorithm, called Fast Incremental Gaussian Mixture Network (FIGMN), was employed as a sample-efficient function approximator for the state space of continuous reinforcement learning tasks, which, combined with linear Q-learning, results in competitive performance. Then, this same function approximator was employed to model the joint state and Q-values space, all in a single FIGMN, resulting in a concise and data-efficient algorithm, i.e., a reinforcement learning algorithm that learns from very few interactions with the environment. A single episode is enough to learn the investigated tasks in most trials. Results are analysed in order to explain the properties of the obtained algorithm, and it is observed that the use of the FIGMN function approximator brings some important advantages to reinforcement learning in relation to conventional neural networks.

Este trabalho introduz novos algoritmos de redes neurais para o processamento online de padrões espaço-temporais, estendendo o algoritmo Incremental Gaussian Mixture Network (IGMN). O algoritmo IGMN é uma rede neural online incremental que aprende a partir de uma única passada através de dados por meio de uma versão incremental do algoritmo Expectation-Maximization (EM) combinado com regressão localmente ponderada (Locally Weighted Regression, LWR). Quatro abordagens diferentes são usadas para dar capacidade de processamento temporal para o algoritmo IGMN: linhas de atraso (Time-Delay IGMN), uma camada de reservoir (Echo-State IGMN), média móvel exponencial do vetor de entrada reconstruído (Merge IGMN) e auto-referência (Recursive IGMN). Isso resulta em algoritmos que são online, incrementais, agressivos e têm capacidades temporais e, portanto, são adequados para tarefas com memória ou estados internos desconhecidos, caracterizados por fluxo contínuo ininterrupto de dados, e que exigem operação perpétua provendo previsões sem etapas separadas para aprendizado e execução. Os algoritmos propostos são comparados a outras redes neurais espaço-temporais em 8 tarefas de previsão de séries temporais. Dois deles mostram desempenhos satisfatórios, em geral, superando as abordagens existentes. Uma melhoria geral para o algoritmo IGMN também é descrita, eliminando um dos parâmetros ajustáveis manualmente e provendo melhores resultados.<br>This work introduces novel neural networks algorithms for online spatio-temporal pattern processing by extending the Incremental Gaussian Mixture Network (IGMN). The IGMN algorithm is an online incremental neural network that learns from a single scan through data by means of an incremental version of the Expectation-Maximization (EM) algorithm combined with locally weighted regression (LWR). Four different approaches are used to give temporal processing capabilities to the IGMN algorithm: time-delay lines (Time-Delay IGMN), a reservoir layer (Echo-State IGMN), exponential moving average of reconstructed input vector (Merge IGMN) and self-referencing (Recursive IGMN). This results in algorithms that are online, incremental, aggressive and have temporal capabilities, and therefore are suitable for tasks with memory or unknown internal states, characterized by continuous non-stopping data-flows, and that require life-long learning while operating and giving predictions without separated stages. The proposed algorithms are compared to other spatio-temporal neural networks in 8 time-series prediction tasks. Two of them show satisfactory performances, generally improving upon existing approaches. A general enhancement for the IGMN algorithm is also described, eliminating one of the algorithm’s manually tunable parameters and giving better results.

Over the last few years autonomous driving had an increasingly strong impact on the automotive industry. This created an increased need for artificial intelligence algo- rithms which allow for computers to make human-like decisions. However, a compro- mise between the computational power drawn by these algorithms and their subsequent performance must be found to fulfil production requirements. In this thesis incremental deep learning strategies are used for the control of a mobile robot such as a four wheel steering vehicle. This strategy is similar to the human approach of learning. In many small steps the vehicle learns to achieve a specific goal. The usage of incremental training leads to growing knowledge-base within the system. It also provides the opportunity to use older training achievements to improve the system, when more training data is available. To demonstrate the capabilities of such an algorithm, two different models have been formulated. First, a more simple model with counter wheel steering, and second, a more complex, nonlinear model with independent steering. These two models are trained incrementally to follow different types of trajectories. Therefore an algorithm was established to generate useful initial points. The incremental steps allow the robot to be positioned further and further away from the desired trajectory in the environ- ment. Afterwards, the effects of different trajectory types on model behaviour are investigated by over one thousand simulation runs. To do this, path planning for straight lines and circles are introduced. This work demonstrates that even simulations with simple network structures can have high performance.

<p>This Masters thesis gives a thorough description of a study carried out in the Self-Organizing Systems group at the NTNU. Much {AI research in the later years has moved towards increased use of representationless strategies such as simulated neural networks. One technique for creating such networks is to evolve them using simulated Darwinian evolution. This is a powerful technique, but it is often limited by the computer resources available. One way to speed up evolution, is to focus the evolutionary search on a more narrow range of solutions. It is for example possible to favor evolution of a specific ``species'' by initializing the search with a specialized set of genes. A disadvantage of doing this is of course that many other solutions (or ``species'') are disregarded so that good solutions in theory may be lost. It is therefore necessary to find focusing strategies that are generally applicable and (with a high probability) only disregards solutions that are considered unimportant. Three different ways of focusing evolutionary search for cognitive behaviours are merged and evaluated in this thesis: On a macro level, incremental evolution is applied to partition the evolutionary search. On a micro level, specific properties of the chosen neural network model (CTRNNs) are exploited. The two properties are seeding initial populations with center-crossing neural networks and/or bifurcative neurons. The techniques are compared to standard, naive, evolutionary searches by applying them to the evolution of simulated neural networks for the walking and control of a six-legged mobile robot. A problem simple enough to be satisfactorily understood, but complex enough to be a challenge for a traditional evolutionary search.</p>