abstract: Anticipatory planning of digit positions and forces is critical for successful dexterous object manipulation. Anticipatory (feedforward) planning bypasses the inherent delays in reflex responses and sensorimotor integration associated with reactive (feedback) control. It has been suggested that feedforward and feedback strategies can be distinguished based on the profile of grip and load force rates during the period between initial contact with the object and object lift. However, this has not been validated in tasks that do not constrain digit placement. The purposes of this thesis were (1) to validate the hypothesis that force rate profiles are indicative of the control strategy used for object manipulation and (2) to test this hypothesis by comparing manipulation tasks performed with and without digit placement constraints. The first objective comprised two studies. In the first study an additional light or heavy mass was added to the base of the object. In the second study a mass was added, altering the object's center of mass (CM) location. In each experiment digit force rates were calculated between the times of initial digit contact and object lift. Digit force rates were fit to a Gaussian bell curve and the goodness of fit was compared across predictable and unpredictable mass and CM conditions. For both experiments, a predictable object mass and CM elicited bell shaped force rate profiles, indicative of feedforward control. For the second objective, a comparison of performance between subjects who performed the grasp task with either constrained or unconstrained digit contact locations was conducted. When digit location was unconstrained and CM was predictable, force rates were well fit to a bell shaped curve. However, the goodness of fit of the force rate profiles to the bell shaped curve was weaker for the constrained than the unconstrained digit placement condition. These findings seem to indicate that brain can generate an appropriate feedforward control strategy even when digit placement is unconstrained and an infinite combination of digit placement and force solutions exists to lift the object successfully. Future work is needed that investigates the role digit positioning and tactile feedback has on anticipatory control of object manipulation. / Dissertation/Thesis / M.S. Bioengineering 2011

Submitted by Automação e Estatística (sst@bczm.ufrn.br) on 2018-05-03T22:57:21Z No. of bitstreams: 1 DanielGarciaTeixeira_DISSERT.pdf: 2395648 bytes, checksum: d18ab806f8ec064f97c3987cacc1c194 (MD5) / Approved for entry into archive by Arlan Eloi Leite Silva (eloihistoriador@yahoo.com.br) on 2018-05-14T21:12:24Z (GMT) No. of bitstreams: 1 DanielGarciaTeixeira_DISSERT.pdf: 2395648 bytes, checksum: d18ab806f8ec064f97c3987cacc1c194 (MD5) / Made available in DSpace on 2018-05-14T21:12:24Z (GMT). No. of bitstreams: 1 DanielGarciaTeixeira_DISSERT.pdf: 2395648 bytes, checksum: d18ab806f8ec064f97c3987cacc1c194 (MD5) Previous issue date: 2018-03-29 / A oscilação gama está presente em diversas áreas do cérebro, como no hipocampo, desempenhando um importante mecanismo para o funcionamento da memória. Encontramos diversos modelos capazes de explicar a geração das oscilações gama e explicam suas duas funcionalidades, agrupamento sincronizado temporalmente das sinapses dos neurônios e a de selecionar quais neurônios devem disparar em cada ciclo deste sincronismo. Funcionalidades estas que imprimem um caráter computacional do processamento neural a este sistema, como a separação de padrões e a formação de assembleias neurais. Porém, a análise destes modelos existentes demonstra ser muito sensível às variações das atividades cerebrais, sendo fortemente afetados por variações nas suas camadas de entrada, de modo a aparentar não possuir uma boa robustez, gerando muita variação de sua frequência de saída, assim como na competitividade entre estes neurônios. Entretanto, ao se considerar uma importante parte do circuito biológico não considerada em trabalhos anteriores, uma rede de inibição alimentada à frente nos possibilitou a criação de um novo modelo. Baseando-nos no modelo de neurônio de Izhikevich, geramos um novo modelo com uma maior estabilidade em sua saída às variações na camada de entrada, bem como um custo computacional reduzido e proximidade do modelo biológico. Em posse deste novo modelo, será possível criar redes neurais com maior capacidade de neurônios, com custo computacional reduzido, além da possibilidade de análise do comportamento individual em cada neurônio do modelo. / Gamma oscillation is present in several areas of the brain, such as the hippocampus, playing an important mechanism for memory functioning. We found several models capable of explaining the generation of the gamma oscillations and explain their two functionalities, that of synchronously grouping the synapses of the neurons and of selecting which neurons must trigger in each cycle of this synchronism. These functionalities impart a computational character of neural processing to this system, such as the separation of patterns and the formation of neural assemblies. However, the analysis of these existent models shows to be very sensitive to the variations of the cerebral activities, being strongly affected by variations and their layers of entrance, in order to appear not to have a good robustness, generating much variation of their frequency of exit, as in between these neurons. However, when considering an important part of the biological circuit not considered in previous studies, a fed-in inhibition network enabled us to create a new model. Based on the Izhikevich neuron model, we generated a new model with greater robustness to the variations in the input layer, as well as a reduced computational cost and proximity of the biological model. In the possession of this new model, it will be possible to create neural networks with greater capacity of neurons, with reduced computational cost, besides the possibility of analyzing the individual behavior in each neuron of the model.

Syftet med arbetet är att simulera tävlingar på medeldistans mellan löpare med en strategi att vinna och undvika muskeltrötthet. Löparna ses som agenter vars strategi realiseras med ett artificiellt neuronnät (ANN) som med sensorer, avstånd till mål och agentens trötthet beräknar bidragande kraft och styrriktning. Agentens ANN tränas med en belöningsbaserad inlärning baserad på genetiska algoritmer och trötthetsalgoritmen är en uppskattning av hur mjölksyra påverkar muskeltrötthet. Resultaten visar att av alla agenter som utvecklats för tävling mot klockan i s.k. time trial har alla haft samma strategi och hittat samma ideala kraft för att minimera tiden. Utvecklingen av agenter för simulation av flera agenter samtidigt har varit mer komplicerad eftersom agenterna påverkar varandra och agenternas strategi har varit olika. Multiagenterna blev också mindre robusta än singelagenterna men utvecklade beteenden som påminner om en realistisk tävling i medeldistanslöpning.

Detta arbete handlar om beteendeinriktad biometri och artificiella neuronnät av typen feedforward och hur de tillsammans kan användas för att verifiera användare. Det har av tidigare arbete bekräftats att det är möjligt att verifiera användare, men tidigare resultat har däremot inte utfört tester med avseende på avvikelser i data (beteende) och dess inverkan på verifieringen. Det är detta som utgör det huvudsakliga målet för detta arbete, nämligen att undersöka hur avvikelser i data påverkar verifiering och utifrån det också undersöka neuronnätens noggrannhet vid verifiering.

Arbetet som presenteras i den här rapporten handlar om ansiktsigenkänning från webbkameror med hjälp av principal component analysis samt artificiella neurala nätverk av typen feedforward. Arbetet förbättrar tekniken med hjälp av filterbaserade metoder som bland annat används inom ansiktsdetektering. Dessa filter bygger på att skicka med redundant data av delregioner av ansiktet.

Detta arbete utvärderar ANN-arkitekturerna feedforward, Elman och ESCN då de används för att styra en bil i en enkel 2D-simulering. Nätverken tränas av en evolutionär algoritm som använder nätverkens vikter som genom för dess individer. Syftet med arbetet är att se om arkitekturerna presterar olika bra. Simuleringens komplexitet, i form av halka och sladd, samt banans svårighetsgrad varieras för att se vilka arkitekturer som klarar vilka komplexa problem bäst och var de eventuellt brister. Ett program utvecklades som testade de olika fallen och resultatet visade att Elman presterade sämst, speciellt då komplexiteten ökade, och ESCN presterade lite bättre än feedforward. Varför Elman presterade sämre fick inget svar i detta arbete, och ESCN använde sitt minne på ett sätt som skulle kunna vara värt att titta vidare på. Framtida arbete skulle kunna vara att ta reda på orsakerna till de ovanliga beteendena som uppstod samt att genomföra mer utförliga tester.

Wastewater compose an environmental risk as it contains high levels of nutrients, including phosphorus. Wastewater treatment plants (WWTPs) reduce phosphorus by using coagulants that precipitate soluble phosphate into metal phosphate, which is separated by settling. Coagulant flow is regulated by a control strategy, typically feedforward or feedback control. Feedforward is based on incoming wastewater disturbances whereas feedback control uses outgoing process values. Incoming phosphate is hard to measure and can be estimated using soft sensors. Modelling control strategies can help decide which strategy that is most suitable. Models describing phosphorus removal are Activated Sludge Model, ASM2d, and primary clarifier model. ASM2d models phosphorus precipitation and the primary clarifier model settling of particles. Tivoli WWTP faces challenges to reach effluent requirements of phosphorus. The wastewater flows through an equalisation tank, Regnbågen, before being pumped to Tivoli. Particulate matter settles in Regnbågen, which is removed by reducing the water level in Regnbågen. This rapidly increases incoming particulate load to Tivoli.Tivoli’s current control strategy is feedforward proportional to suspended solids. It is suspected, that this strategy overdose coagulant during the emptying of Regnbågen. The purpose of this thesis was to find the optimal control strategy for phosphorus precipitation at Tivoli WWTP, by using a model-based approach. Control strategies modelled are; feedforward, feedback and these two control strategies combined. Additional issues resolved are 1) calibration of a model that predicts the effect of chemical dosage on effluent phosphorus concentration from the primary clarifier, 2) calibrationof a soft sensor, 3) determination of which control strategy that is most suitable. ASM2d and a primary clarifier model were used to create a model describing chemical phosphorus removal. The calibration matches measured phosphate concentration, but underestimate peaks. The primary clarifier model was calibrated by minimising load differences for phosphate and total suspended solids, and was calibrated satisfyingly. A simplified soft sensor was constructed, described by a linear relationship between phosphate and pH. Three disturbances for feedforward control were analysed; measured phosphate, the soft sensors estimation of phosphate and Tivoli’s current controlstrategy. The optimal control strategy was found through a multi-criteria analysis. The optimal control strategy is the combined control strategy, when feedforward is proportional to incoming measured phosphate. The performance of all analysed feedforward disturbances were significantly improved when combined with feedback control. Furthermore, consequential errors are distinct when the soft sensor miss-predictincoming phosphate concentration. If the phosphate concentration cannot be correctly measured/estimated, feedback control alone has the best performance.

The phenomenon of overfitting, where a feed-forward neural network (FFNN) over trains on training data at the cost of generalisation accuracy is known to be speci c to the training algorithm used. This study investigates over tting within the context of particle swarm optimised (PSO) FFNNs. Two of the most widely used PSO algorithms are compared in terms of FFNN accuracy and a description of the over tting behaviour is established. Each of the PSO components are in turn investigated to determine their e ect on FFNN over tting. A study of the maximum velocity (Vmax) parameter is performed and it is found that smaller Vmax values are optimal for FFNN training. The analysis is extended to the inertia and acceleration coe cient parameters, where it is shown that speci c interactions among the parameters have a dominant e ect on the resultant FFNN accuracy and may be used to reduce over tting. Further, the signi cant e ect of the swarm size on network accuracy is also shown, with a critical range being identi ed for the swarm size for e ective training. The study is concluded with an investigation into the e ect of the di erent activation functions. Given strong empirical evidence, an hypothesis is made that stating the gradient of the activation function signi cantly a ects the convergence of the PSO. Lastly, the PSO is shown to be a very effective algorithm for the training of self-adaptive FFNNs, capable of learning from unscaled data. / Dissertation (MSc)--University of Pretoria, 2014. / tm2015 / Computer Science / MSc / Unrestricted

An optimal stochastic feedforward-feedback control scheme is implemented on a heat exchanger-stirred tank system using an on-line minicomputer. Because variations in the measured disturbance variable have an effect on the output controlled variable before compensating action can become effective, the feedforward action must be predictive in nature. Statistical time series models are used to model both the measured disturbance and the unobserved disturbances in the system. These stochastic disturbance models and the transfer function models for the process are identified, fitted and checked using statistical model building procedures on a set of data collected on-line using the minicomputer. The predictive feedforward-feedback controller is derived from these models. The performance of the control scheme is compared with that of a pure feedback control scheme and the actual performances are shown to conform well to the theory. / Thesis / Master of Engineering (ME)

Using the data collected during the hyperparameter tuning for Google's Dependency Parser Parsey McParseface, Feedforward neural networks and the correlation between its hyperparameter during the networks training are explained and analysed in depth.:1 Introduction to Neural Networks 4 1.1 History of AI 4 1.2 The role of Neural Networks in AI Research 6 1.2.1 Artificial Intelligence 6 1.2.2 Machine Learning 6 1.2.3 Neural Network 8 1.3 Structure of Neural Networks 8 1.3.1 Biology Analogy of Artificial Neural Networks 9 1.3.2 Architecture of Artificial Neural Networks 9 1.3.3 Biological Model of Nodes – Neurons 11 1.3.4 Structure of Artificial Neurons 12 1.4 Training a Neural Network 21 1.4.1 Data 21 1.4.2 Hyperparameters 22 1.4.3 Training process 26 1.4.4 Overfitting 27 2 Natural Language Processing (NLP) 29 2.1 Data Preparation 29 2.1.1 Text Preprocessing 29 2.1.2 Part-of-Speech Tagging 30 2.2 Dependency Parsing 31 2.2.1 Dependency Grammar 31 2.2.2 Dependency Parsing Rule-Based & Data-Driven Approach 33 2.2.3 Syntactic Parser 33 2.3 Parsey McParseface 34 2.3.1 SyntaxNet 34 2.3.2 Corpus 34 2.3.3 Architecture 34 2.3.4 Improvements to the Feed Forward Neural Network 38 3 Training of Parsey’s Cousins 41 3.1 Training a Model 41 3.1.1 Building the Framework 41 3.1.2 Corpus 41 3.1.3 Training Process 43 3.1.4 Settings for the Training 44 3.2 Results and Analysis 46 3.2.1 Results from Google’s Models 46 3.2.2 Effect of Hyperparameter 47 4 Conclusion 63 5 Bibliography 65 6 Appendix 74