Thesis (M.S.)--Ohio University, August, 1991. / Title from PDF t.p.

A mathematical model for goal oriented learning in a network of neuron-like elements was developed. Using a mouse/goal box analogy, a simulation of a network with four elements was programmed in Turbo Pascal, Version 4.0 (Borland International) to test the model. Each location in the network corresponded to a particular network input. The output of the network consisted of one of four behaviors: forward, backward, left or right. The network successfully learned sequences of up to six movements in increasingly complex mazes. / Master of Science

Thesis (MScEng)--Stellenbosch University, 2002. / ENGLISH ABSTRACT: Flotation is notorious for its susceptibility to process upsets and consequently its poor performance, making successful flotation control systems an elusive goal. The control of industrial flotation plants is often based en the visual appearance of the froth phase, and depends to a large extent on the experience and ability of a human operator. Machine vision systems provide a novel solution to several of the problems encountered in conventional flotation systems for monitoring and control. The rapid development in computer VISIon, computational resources and artificial intelligence and the integration of these technologies are creating new possibilities in the design and implementation of commercial machine vision systems for the monitoring and control of flotation plants. Current machine vision systems are available but not without their shortcomings. These systems cannot deal with fine froths where the bubbles are very small due to the segmentation techniques employed by them. These segmentation techniques are cumbersome and computationally expensive making them slow in real time operation. The approach followed in this work uses neural networks to solve the problems mentioned above. Neural networks are able to extract information from images of the froth phase without regard to the type and structure of the froth. The parallel processing capability of neural networks, ease of implementation and the advantages of supervised or unsupervised training of neural networks make them potentially suited for real-time industrial machine vision systems. In principle, neural network models can be implemented in an adaptive manner, so that changes in the characteristics of processes are taken into account. This work documents the development of linear and non-linear principal component models, which can be used in a real-time machine vision system for the monitoring, and control of froth flotation systems. Features from froth images of flotation processes were extracted via linear and non-linear principal component analysis. Conventional linear principal component analysis and three layer autoassociative neural networks were used in the extraction of linear principal components from froth images. Non-linear principal components were extracted from froth images by a three and five layer autoassociative neural network, as well as localised principal component analysis based on k-means clustering. Three principal components were extracted for each image. The correlation coefficient was used as a measure of the amount of variance captured by each principal component. The principal components were used to classify the froth images. A probabilistic neural network and a feedforward neural network classifier were developed for the classification of the froth images. Multivariate statistical process control models were developed using the linear and non-linear principal component models. Hotellings T2 statistic and the squared prediction error based on linear and non-linear principal component models were used in the development of multivariate control charts. It was found that the first three features extracted with autoassociative neural networks were able to capture more variance in froth images than conventional linear principal components, the features extracted by the five layer autoassociative neural networks were able to classify froth images more accurately than features extracted by conventional linear principal component analysis and three layer autoassociative neural networks. As applied, localised principal component analysis proved to be ineffective, owing to difficulties with the clustering of the high dimensional image data. Finally the use of multivariate statistical process control models to detect deviations from normal plant operations are discussed and it is shown that Hotellings T2 and squared prediction error control charts are able to clearly identify non-conforming plant behaviour. / AFRIKAANSE OPSOMMING: Flottasie is berug daarvoor dat dit vatbaar vir prosesversteurings is en daarom dikwels nie na wense presteer nie. Suksesvolle flottasiebeheerstelsels bly steeds 'n ontwykende doelwit. Die beheer van nywerheidsflottasie-aanlegte word dikwels gebaseer op die visuele voorkoms van die skuimfase en hang tot 'n groot mate af van die ervaring en vaardighede van die menslike operateur. Masjienvisiestelsels voorsien 'n vindingryke oplossing tot verskeie van die probleme wat voorkom by konvensionele flottasiestelsels ten opsigte van monitering en beheer. Die vinnige ontwikkeling van rekenaarbeheerde visie, rekenaarverwante hulpbronne en kunsmatige intelligensie, asook die integrasie van hierdie tegnologieë, skep nuwe moontlikhede in die ontwerp en inwerkingstelling van kommersiële masjienvisiestelsels om flottasie-aanlegte te monitor en te beheer. Huidige masjienvisiestelsels is wel beskikbaar, maar is nie sonder tekortkominge nie. Hierdie stelsels kan nie fyn skuim hanteer nie, waar die borreltjies baie klein is as gevolg van die segmentasietegnieke wat hulle aanwend. Hierdie segmentasietegnieke is omslagtig en rekenaargesproke duur, wat veroorsaak dat dit stadig in reële tyd-aanwendings is. Die benadering wat in hierdie werk gevolg is, wend neurale netwerke aan om die bovermelde probleme op te los. Neurale netwerke is instaat om inligting te onttrek uit beelde van die skuimfase sonder om ag te slaan op die tipe en struktuur van die skuim. Die parallelle prosesseringsvermoëns van neurale netwerke, die gemak van implementering en die voordele van die opleiding van neurale netwerke met of sonder toesig maak hulle potensieel nuttig as reële tydverwante industriële masjienvisiestelsels. In beginsel kan neurale netwerke op 'n aanpassende wyse geïmplementeer word, sodat veranderinge in die kenmerke van die prosesse deurlopend in aanmerking geneem word. Kenmerke van die beelde van die skuim tydens die flottasieproses is verkry by wyse van lineêre en nie-lineêre hootkomponentsanalise. Konvensionele lineêre hoofkomponentsanalise en drie-laag outo-assosiatiewe neurale netwerke is gebruik in die onttrekking van lineêre hoofkomponente uit die beelde van die skuim. Nie-lineêre hoofkomponente is uit die beelde van die skuim onttrek by wyse van 'n drie- en vyf-laag outo-assosiatiewe neurale netwerk, asook deur 'n gelokaliseerde hoofkomponentsanalise wat op k-gemiddelde trosanalise gebaseer is. Drie hoofkomponente is vir elke beeld onttrek. Die korrelasiekoëffisiënt is gebruik as 'n maatstaf van die afwyking wat deur elke hoofkomponent aangetoon is. Die hoofkomponente is gebruik om die beelde van die skuim te klassifiseer. 'n Probalistiese neurale netwerk en 'n voorwaarts voerende neurale netwerk is vir die klassifisering van die beelde van die skuim ontwerp. Multiveranderlike statistiese prosesbeheermodelle is ontwerp met die gebruik van die lineêre en nie-lineêre hoofkomponentmodelle. Hotelling se T2 statistiek en die gekwadreerde voorspellingsfout, gebaseer op lineêre en nie-lineêre hoofkomponentmodelle, is gebruik in die ontwikkeling van multiveranderlike kontrolekaarte. Dit is gevind dat die eerste drie eienskappe wat met behulp van die outo-assosiatiewe neurale netwerke onttrek is, instaat was om meer variansie by beelde van skuim vas te vang as konvensionele lineêre hoofkomponente. Die eienskappe wat deur die vyf-laag outo-assosiatiewe neurale netwerke onttrek is, was instaat om beelde van skuim akkurater te klassifiseer as daardie eienskappe wat by wyse van konvensionele lineêre hoofkomponentanalalise en drie-laag outo-assosiatiewe neurale netwerke onttrek is. Soos toegepas, het dit geblyk dat gelokaliseerde hoofkomponentsanalise nie effektief is nie, as gevolg van die probleme rondom die trosanalise van die hoë-dimensionele beelddata. Laastens word die aanwending van multiveranderlike statistiese prosesbeheermodelle, om afwykings in normale aanlegoperasies op te spoor, bespreek. Dit word aangetoon dat Hotelling se T2 statistiek en gekwadreerdevoorspellingsfoutbeheerkaarte instaat is om afwykende aanlegwerksverrigting duidelik aan te dui.

No description available.

Thesis (M.Phil.)--Hong Kong University of Science and Technology, 2005. / Includes bibliographical references (leaves 68-71). Also available in electronic version. Access restricted to campus users.

Thesis (M.S.)--Virginia Polytechnic Institute and State University, 1993. / Vita. Abstract. Includes bibliographical references (leaves 134-135). Also available via the Internet.

Much of our interaction with the environment is physical. We use our bodies for nonverbal expression or to augment or emphasize verbal communication. In other cases we use our bodies to execute tasks such as walking or picking up an object. A human observer can easily recognise these activities. For example, it is the job of a security officer in a supermarket to observe people and check that articles are not stolen. If a person does steal, the security officer recognises the act and takes appropriate action. The problem addressed in this study is the automatic recognition of human gestures by means of video image analysis. For this purpose a computer-based system with similar recognition capabilities as a human observer is investigated. The system uses cameras that correspond to the eyes and algorithms that resemble abilities of the human visual system. Automatic gesture recognition is a complex problem and the focus here is to develop algorithms that will solve a subset of the problem. This involves the recognition of simple gestures such as walking and waving of arms. The approach taken in this dissertation is to represent body shape in camera images with a simple model called a bounding box. This model has the appearance of a rect¬angle that encapsulates the extremities of the human body and resembles the coarse structure of body shape. From a representation point of view, the model is an abstrac¬tion of body pose. A gesture consists of a sequence of poses. By employing pattern recognition techniques, a sequence of pose abstractions is recognised as a gesture. Various aspects of the bounding box model are explored in this study. Perception experiments are conducted to gain a conceptual understanding of the behaviour of the model. Other aspects include investigation of two- and three-dimensional spatial representations of the model with a neural network classifier as well as the model's temporal properties through the use of hidden Markov models. These aspects are tested using gesture recognition systems implemented for this purpose. The gesture vocabularies of these systems range from four to ten gestures, while recognition rates vary from 84.7% to 96.3%. / Dissertation (MEng (Computer Engineering))--University of Pretoria, 2006. / Electrical, Electronic and Computer Engineering / unrestricted

This thesis describes an approach for accomplishing a pattern recognition task using conceptual graph theory and neural networks (NNs). The set of patterns to be recognized are the capital letters of six different fonts of the English alphabet, plus two shifted and six rotated versions of each. The letters are represented to the neural network on a 16x16 input grid (256 "sensor lines"). A standard classification encoding for such patterns is to use a 26-bit vector (26 lines at the NN's output), one bit corresponding to each letter. Experiments with such an encoding yielded results with poor generalization capability. A new encoding scheme was developed, based on the conceptual graph formalism. This entailed designing a set of concepts and a set of associated relations appropriate to the upper case letters of the English alphabet. From these, the following were developed: a conceptual graph representation for each letter, a connection matrix for each, and finally, a C-vector and an R-vector representation for each. The latter were used as the output encoding (21 bits) of the NN pattern recognizer. A feed-forward neural network with 256 inputs, 21 outputs, and 2 hidden layers was trained using the back-propagation- of-error algorithm. Results were significantly better than with the more standard. encoding. Generalization on fonts improved from 74% to 96%, generalization on rotations improved from 83% to 94%, and finally, generalization on shifts improved from 2% to 93%.

Brains display very high-level parallel computation, fault-tolerance, and adaptability, all of which are what we struggle to recreate in engineered systems. The neurocomputer (an organic computer built from living neurons) seems possible and may lead to a new generation of computing device that can operate in a brain-like manner. Cultured neuronal networks on multi-electrode arrays (MEAs) are one of the best candidates for the neurocomputer for their controllability, accessibility, flexibility, and the ability to self-organize. I explored the possibility of the neurocomputer by studying whether we can show goal-directed learning, one of the most fascinating behavior of brains, in cultured networks. Inspired by the brain, which needs to be embodied in some way and interact with its surroundings in order to give a purpose to its activities, we have developed tools for closing the sensory-motor loop between a cultured network and a robot or an artificial animal (an animat), termed a ¡§hybrot¡¨. In order to efficiently find an effective closed-loop design among infinite potential options, I constructed a biologically-inspired simulated network. By using this simulated network, I designed: (1) a statistic that can effectively and efficiently decode network functional plasticity, and (2) feedback stimulations and an adaptive training algorithm to encode sensory information and to direct network plasticity. By closing the sensory-motor loop with these decoding and encoding designs, we successfully demonstrated a simple adaptive goal-directed behavior: learning to move in a user-defined direction, and further showed that multiple tasks could be learned simultaneously. These results suggest that even though a cultured network lacks the 3-D structure of the brain, it still can be functionally shaped and show meaningful behavior. To our knowledge, this is the first demonstration of goal-directed learning in embodied cultured networks. Extending from these findings, I further proposed a research plan to optimize closed-loop designs for evaluating the maximal learning capacity (or even true intelligence) of the cultured network. Knowledge gained from effective closed-loop designs provides insights about learning and memory in the nervous system, which could influence the design of neurocomputers, future artificial neural networks, and more effective neuroprosthetics.

Artificial neural networks (ANN) can detect complex non-linear relationships between independent and dependent variables. Properly trained ANNs have repeatedly demonstrated superior predictive accuracy to other predictive technologies when applied to non-linear systems. Currently there are no studies that have been carried out on predicting survival of leukaemia patients at all. The neural network prediction method adopted in this study aims to provide a robust and accurate method for predicting survival of leukaemia patients for both censored and uncensored patient data. The aim of this research was also to find out the effectiveness of neural networks in modelling leukaemia prognosis and to determine the factors that have the most influence. There is ongoing research into finding ways and means of extending the life span of diseased patients. There is great interest in identifying factors that will yield better predictions of survival for terminally ill leukaemia patients. Prognostic factors generally differ with the treatment of leukaemia. Clinicians face the problem of how to choose the appropriate treatment regime, therefore an analysis of prognostic factors that predict success or failure may identify patients who require an alternative approach of specialist or targeted treatment. Being able to predict an individual patient’s prognosis will enable clinicians to categorise them into the relevant high and low risk treatment groups for conventional treatment or allow for the patients to be incorporated into specialised treatment schedules and clinical trials if available. In this study there is believed to be relationship that exists between the results gained on diagnosis and the period of survival. A patient’s health status is dependent on various symptoms and the complexity of the medical condition is dependent on an individual’s biological system. This complexity allows for the application of artificial neural networks (ANN) in predicting outcomes in medical application, especially in prognosis prediction and survival rate. This thesis contains contributions to the development of neural network models for survival analysis of leukaemia patients. The feed forward back propagation algorithm (BPA) modified to the gradient descent BPA was identified for the training and building of the neural network for predicting survival of leukaemia patients. The prognostic factors that affect survival have also been determined by the neural networks. The comparisons of models were based on using combined groups of leukaemia patients and comparing them with individual groups of the sub-types of leukaemia, i.e. acute lymphoid leukaemia (ALL), acute myeloid leukaemia (AML), chronic myeloid leukaemia (CLL) and chronic myeloid leukaemia (CML). A combination of 38 variables was used in the development of the neural networks. The variables were age, race, sex, gender, and results of full blood counts, differential tests and flow cytometry. The survival period of patients was based on the diagnosis date and the date of treatment. Those patients who status of mortality was known as of October 2008 were considered to be uncensored and were used for the 2-year and 3-year case studies. The patients with unknown mortality were considered as censored patients and used for the censored case study. The patient data was processed into a coded system and used to build the neural networks for each data set. The choice of patient groups used for the model building was prompted by the availability of uncensored data for analysis. For the group of combined leukaemia patients and the sub-group CML-CLL, it is recommended that the 2-year neural network model be used. The main prognostic factors affecting leukaemia survival were found to be the patient’s age, the mean haemoglobin concentration, % neutrophils and the markers CD13, CD20 and CD56. The race group, platelet count, % monocytes and the markers CD3, CD4, CD34 and LC lambda were found to significantly affect the CML-CLL group of patients. For the ALL and AML groups the 3-year neural network models were favoured. Prognostic factors for the survival of ALL patients were their age, the mean corpuscular haemoglobin concentration, % blasts and the markers CD8 and CD22. For the AML group the important prognostic factors were the patient’s age, the mean corpuscular haemoglobin concentration, the % neutrophils, % lymphocytes, and the markers CD7 and CD34. / Thesis (Ph.D.)-University of KwaZulu-Natal, Durban, 2009.