Undergraduate biomechanics is classically viewed as one of the most difficult courses included in kinesiology programs, often leading to poor student performance and attitudes. By adjusting the interactions students have with course material, it may be possible to positively impact student outcomes. Past work has shown that interactive learning episodes can positively impact student attitudes toward difficult course content, as well as improve student performance variables (Catena & Carbonneau, n.d.; Moreno & Mayer, 2007; Pandy, Petrosino, Austin, & Barr, 2004; Zhang, Zhou, Briggs, & Nunamaker, 2005). In the present study, I investigated the effectiveness of interactive, exploratory based learning episodes in undergraduate biomechanics laboratory sessions. Episodes consisted of a brief introduction of the laboratory topic, which was consistent across groups, followed immediately by a pre- laboratory assessment. Students then completed the laboratory, which either included exploration in interactive computer applications or still images of the applications displaying the necessary information for completion. Intervention sessions utilized custom interactive computer applications where students were prompted to explore course concepts centered around reciprocal relationships between variables specific to each laboratory topic. Student performance was collected and assessed for Work Loop Muscle Mechanics and EMG signal processing laboratory topics at two independent instances. For both learning topics, intervention and control groups both, improved their scores between pre- and post-laboratory assessments indicating learning. In the post-laboratory testing, the intervention group significantly outperformed the control group on the most challenging assessment question (P = 0.005). Adversely, the intervention group achieved significantly lower scores for the simplest signal processing questionnaire item (P <0.001). Although the present study contained mixed results, it supports the utilization of exploratory based learning episodes on typically challenging topics with abstract concepts. Further investigation is needed in order to explore the chronic learning effects of such instructional methods. / Graduate

<p>Spiking Neural Networks (SNNs) have recently attracted significant research interest as the third generation of artificial neural networks that can enable low-power event-driven data analytics. We propose ANN-SNN conversion using “soft re-set” spiking neuron model, referred to as Residual Membrane Potential (RMP) spiking neuron, which retains the “resid- ual” membrane potential above threshold at the firing instants. In addition, we propose a time-based coding scheme, named Temporal-Switch-Coding (TSC), and a corresponding TSC spiking neuron model. Each input image pixel is presented using two spikes with opposite polarity and the timing between the two spiking instants is proportional to the pixel intensity. We demonstrate near loss-less ANN-SNN conversion using RMP neurons for VGG-16, ResNet-20, and ResNet-34 SNNs on challenging datasets including CIFAR-10, CIFAR-100, and ImageNet. With the help of TSC coding, it achieves 7-14.5× less inference latency, and 30-60× fewer addition operations and memory accesses per inference across datasets compared to the state of the art (SOTA) SNN models. In the second part of the thesis, we propose a new type of recurrent neural network (RNN) architecture, named Os- cillatory Fourier Neural Network (O-FNN). We demonstrate that O-FNN is mathematically equivalent to a simplified form of Discrete Fourier Transform applied onto periodical activa- tion. In particular, the computationally intensive back-propagation through time in training is eliminated, leading to faster training while achieving the SOTA inference accuracy in a diverse group of sequential tasks. For instance, applying the proposed model to sentiment analysis on IMDB review dataset reaches 89.4% test accuracy within 5 epochs, accompanied by over 35x reduction in the model size compared to Long Short-Term Memory (LSTM). The proposed novel RNN architecture is well poised for intelligent sequential processing in resource constrained hardware.</p>

There is a large demand for the use of e-learning tools to support student learning, in the form of distance or blended learning. The need for e-learning environment that encourages learners to learn independently or in groups in virtual settings is crucial. Some e-learning environments provide repositories of 'resources'. They neither facilitate a strategy for learning or teaching, nor they guide students through the resources, and tutors in constructing their courses. E-learning environments need to incorporate pedagogical practices which support and allow students to learn by removing any barriers that might inhibit their learning. Therefore, one of the most important aspects in developing e-learning environments is defining appropriate models where technology and pedagogy are integrated. This thesis provides such a framework for developing e-learning applications; it aims to make it easier for tutors to implement their lesson content and engage learners to achieve the course objectives. The proposed model incorporates constructive alignment, assessment and feedback and unlike other e-learning environments guides the tutor to construct lessons and help learners to use effective learning environment. Furthermore, the thesis investigates on how supported learning can help students adapt to the different approaches to learning. The empirical work undertaken investigates the role of constructing a well designed self-assessment and feedback unit within a learning environment.

The purpose of this study is to evaluate the effectiveness of an intervention model designed to enhance practitioners’ biological lens when using a biopsychosocial-spiritual model of holistic assessment and planning. The specific intervention utilized is a course curriculum developed to broaden human service professionals’ (including clinical social work professionals) understanding of attachment theory, neuroscience and trauma informed methods of practice. The course teaches professionals how to apply this knowledge to clinical assessment and intervention planning with youth who have experienced significant trauma in their lives and exhibit problems of conduct. Using an experimental design, participants from a large private mental health organization were asked to evaluate the impact of curriculum on their 1) knowledge of attachment theory, trauma informed practice and neurobiology; 2) attitudes concerning the relevance of trauma-informed practice, the biological perspective and consequence focused models of intervention; and 3) assessment and intervention planning strategies. The curriculum focused its application on youth who have experienced significant levels of trauma and display conduct related behavior problems. Group differences for the workshop intervention group and waitlist control group are discussed. Additionally, a preliminary evaluation of differences between two different intervention groups (participants in the Distance Learning version of the course and the Workshop Seminar version of the course) was conducted.

This project deals with the study of the social learning dynamics of agents in a society. For that we employ techniques from statistical mechanics, machine learning and probability theory. Agents interact in pairs by exchanging for/against opinions about issues using an algorithm constrained by available information. Making use of a maximum entropy analysis one can describe the interacting pair as a dynamics along the gradient of the logarithm of the evidence. This permits introducing energy like quantities and approximate global Hamiltonians. We test different hypothesis having in mind the limitations and advantages of each one. Knowledge of the expected value of the Hamiltonian is relevant information for the state of the society, inducing a canonical distribution by maximum entropy. The results are interpreted with the usual tools from statistical mechanics and thermodynamics. Some of the questions we discuss are: the existence of phase transitions separating ordered and disordered phases depending on the society parameters; how the issue being discussed by the agents influences the outcomes of the discussion, and how this reflects on the overall organization of the group; and the possible different interactions between opposing parties, and to which extent disagreement affects the cohesiveness of the society. / Esse projeto lida com o estudo da dinâmica de aprendizado social de agentes em uma sociedade. Para isso empregamos técnicas de mecânica estatística, aprendizado de máquina e teoria de probabilidades. Agentes interagem em pares trocando opiniões pró/contra questões usando um algoritmo restringido pela informação disponível. Fazendo-se uso de uma análise de máxima entropia, pode-se descrever o par da interação como uma dinâmica ao longo do gradiente do logaritmo da evidência. Isso permite introduzir quantidades similares a energia e Hamiltonianos globais aproximados. Testamos diferentes hipóteses tendo em mente as limitações e as vantagens de cada uma. Conhecimento do valor esperado do Hamiltoniano é informação relevante para o estado da sociedade, induzindo uma distribuição canônica a partir de máxima entropia. Os resultados são interpretados com as ferramentas usuais de mecânica estatística e termodinâmica. Algumas das questões que discutimos são: a existência de transições de fase separando fases ordenada e desordenada dependendo dos parâmetros da sociedade; o como a questão sendo discutida pelos agentes influencia os resultados da discussão, e como isso se reflete na organização do grupo como um todo; e as possíveis diferentes interações entre partidos opostos, e até que ponto o desacordo afeta a coesão da sociedade.

Neste trabalho foi elaborada uma estrutura de modelos probabilísticos simples que pudessem descrever o processo de tomada de decisão de agentes humanos que são confrontados com a tarefa de prever elementos de uma sequência aleatória gerada por uma cadeia de Markov de memória L. Essa estrutura partiu de uma abordagem bayesiana em que o agente infere uma distribuição de probabilidades a partir de uma série de observações da sequência e de suas próprias respostas, considerando que o agente tenha uma memória de tamanho K. Como resultado da abordagem bayesiana, o agente adota uma estratégia ótima que consiste na perseveração na alternativa mais provável dado o histórico das últimas tentativas; por conta disso e de observações experimentais de que humanos tendem a adotar nesse tipo de problema estratégias sub-ótimas, por exemplo a de pareamento de probabilidades (probability matching), foram desenvolvidas variações sobre esse modelo que tentassem descrever mais de perto o comportamento adotado por humanos. Nesse sentido, foram adotadas as variáveis de troca de resposta (possível ação tomada pelo agente) e de recompensa (possível resultado da ação) na formulação do modelo e foram adicionados parâmetros, inspirados em modelos de ação dopaminérgica, que permitissem um desvio da estratégia ótima resultante da abordagem bayesiana. Os modelos construídos nessa estrutura foram simulados computacionalmente para diversos valores dos parâmetros, incluindo as memórias K e L do agente e da cadeia de Markov, respectivamente. Através de análises de correlação, esses resultados foram comparados aos dados experimentais, de um grupo de pesquisa do Instituto de Ciências Biomédicas da USP, referentes a tarefas de tomada de decisão envolvendo pessoas de diversas faixas etárias (de 3 a 73 anos) e cadeias de Markov de memórias 0, 1 e 2. Nessa comparação, concluiu-se que as diferenças entre grupos etários no experimento podem ser explicadas em nossa modelagem através da variação da memória K do agente crianças de até 5 anos mostram um limite K = 1, e as de até 12 anos mostram um limite K = 2 e da variação de um parâmetro de reforço de aprendizado dependendo do grupo e da situação de decisão à qual os indivíduos eram expostos, o valor ajustado desse parâmetro variou de 10% para baixo até 30% para cima do seu valor original de acordo com a abordagem bayesiana. / In this work we developed a simple probabilistic modeling framework that could describe the process of decision making in human agents that are presented with the task of predicting elements of a random sequence generated by a Markov chain with memory L. Such framework arised from a Bayesian approach in which the agent infers a probability distribution from a series of observations on the sequence and on its own answers, and considers that the agent\'s memory has length K. As a result of the Bayesian approach, the agent adopts an optimal strategy that consists in perseveration of the most likely alternative given the history of the last few trials; because of that and of experimental evidence that humans tend, in such kinds of problems, to adopt suboptimal strategies such as probability matching, variations on that model were developed in an attempt to have a closer description of the behavior adopted by humans. In that sense, the `shift\' (possible action taken by the agent on its response) and `reward\' (possible result of the action) variables were adopted in the formulation of the model, and parameters inspired by models of dopaminergic action were added to allow deviation from the optimal strategy that resulted from the Bayesian approach. The models developed in that framework were computationally simulated for many values of the parameters, including the agent\'s and the Markov chain\'s memory lengths K and L respectively. Through correlation analysis these results were compared to experimental data, from a research group from the Biomedical Science Institute at USP, regarding decision making tasks that involved people of various ages (3 to 73 years old) and Markov chains of orders 0, 1 and 2. In this comparison it was concluded that the differences between age groups in the experiment can be explained in our modeling through variation of the agent\'s memory length K children up to 5 years old exhibited a limitation of K = 1, and those up to 12 years old were limited to K = 2 and through variation of a learning reinforcement parameter depending on the group and the decision situation to which the candidates were exposed, the fitted value for that parameter ranged from 10% below to 30% above its original value according to the Bayesian approach.

The coverage of a learning algorithm is the number of concepts that can be learned by that algorithm from samples of a given size for given accuracy and confidence parameters. This thesis begins by asking whether good learning algorithms can be designed by maximizing their coverage. There are three questions raised by this approach: (i) For given sample size and other learning parameters, what is the largest possible coverage that any algorithm can achieve? (ii) Can we design a learning algorithm that attains this optimal coverage? (iii) What is the coverage of existing learning algorithms? This thesis contributes to answering each of these questions. First, we generalize the upper bound on coverage given in [Dietterich 89]. Next, we present two learning algorithms and determine their coverage analytically. The coverage of the first algorithm, Multi-Balls, is shown to be quite close to the upper bound. The coverage of the second algorithm, Large-Ball, turns out to be even better than Multi-Balls in many situations. Third, we considerably improve upon Dietterich's limited experiments for estimating the coverage of existing learning algorithms. We find that the coverage of Large-Ball exceeds the coverage of ID3 [Quinlan 86] and FRINGE [Pagano and Haussler 90] by more than an order of magnitude in most cases. Nevertheless, further analysis of Large-Ball shows that this algorithm is not likely to be of any practical help. Although this algorithm learns many concepts, these do not seem to be very interesting concepts. These results lead us to the conclusion that coverage maximization alone does not appear to yield practically-useful learning algorithms. The results motivate considering the biased-coverage under which different concepts are assigned different weight or importance based on given background assumptions. As an example of the new setting, we consider learning situations where many of the features present in the domain are irrelevant to the concept being learned. These situations are often encountered in practice. For this problem, we define and study the MIN-FEATURES bias in which hypotheses definable using a smaller number of features involved are preferred. We prove a tight bound on the number of examples needed for learning. Our results show that, if the MIN-FEATURES bias is implemented, then the presence of many irrelevant features does not make the learning problem substantially harder in terms of the needed number of examples. The thesis also introduces and evaluates a number of algorithms that implement or approximate the MIN-FEATURES bias. / Graduation date: 1993

Thesis (Ph.D.)--George Mason University, 1999. / Includes bibliographical references (leaves [180]-189).

This project deals with the study of the social learning dynamics of agents in a society. For that we employ techniques from statistical mechanics, machine learning and probability theory. Agents interact in pairs by exchanging for/against opinions about issues using an algorithm constrained by available information. Making use of a maximum entropy analysis one can describe the interacting pair as a dynamics along the gradient of the logarithm of the evidence. This permits introducing energy like quantities and approximate global Hamiltonians. We test different hypothesis having in mind the limitations and advantages of each one. Knowledge of the expected value of the Hamiltonian is relevant information for the state of the society, inducing a canonical distribution by maximum entropy. The results are interpreted with the usual tools from statistical mechanics and thermodynamics. Some of the questions we discuss are: the existence of phase transitions separating ordered and disordered phases depending on the society parameters; how the issue being discussed by the agents influences the outcomes of the discussion, and how this reflects on the overall organization of the group; and the possible different interactions between opposing parties, and to which extent disagreement affects the cohesiveness of the society. / Esse projeto lida com o estudo da dinâmica de aprendizado social de agentes em uma sociedade. Para isso empregamos técnicas de mecânica estatística, aprendizado de máquina e teoria de probabilidades. Agentes interagem em pares trocando opiniões pró/contra questões usando um algoritmo restringido pela informação disponível. Fazendo-se uso de uma análise de máxima entropia, pode-se descrever o par da interação como uma dinâmica ao longo do gradiente do logaritmo da evidência. Isso permite introduzir quantidades similares a energia e Hamiltonianos globais aproximados. Testamos diferentes hipóteses tendo em mente as limitações e as vantagens de cada uma. Conhecimento do valor esperado do Hamiltoniano é informação relevante para o estado da sociedade, induzindo uma distribuição canônica a partir de máxima entropia. Os resultados são interpretados com as ferramentas usuais de mecânica estatística e termodinâmica. Algumas das questões que discutimos são: a existência de transições de fase separando fases ordenada e desordenada dependendo dos parâmetros da sociedade; o como a questão sendo discutida pelos agentes influencia os resultados da discussão, e como isso se reflete na organização do grupo como um todo; e as possíveis diferentes interações entre partidos opostos, e até que ponto o desacordo afeta a coesão da sociedade.

Neste trabalho foi elaborada uma estrutura de modelos probabilísticos simples que pudessem descrever o processo de tomada de decisão de agentes humanos que são confrontados com a tarefa de prever elementos de uma sequência aleatória gerada por uma cadeia de Markov de memória L. Essa estrutura partiu de uma abordagem bayesiana em que o agente infere uma distribuição de probabilidades a partir de uma série de observações da sequência e de suas próprias respostas, considerando que o agente tenha uma memória de tamanho K. Como resultado da abordagem bayesiana, o agente adota uma estratégia ótima que consiste na perseveração na alternativa mais provável dado o histórico das últimas tentativas; por conta disso e de observações experimentais de que humanos tendem a adotar nesse tipo de problema estratégias sub-ótimas, por exemplo a de pareamento de probabilidades (probability matching), foram desenvolvidas variações sobre esse modelo que tentassem descrever mais de perto o comportamento adotado por humanos. Nesse sentido, foram adotadas as variáveis de troca de resposta (possível ação tomada pelo agente) e de recompensa (possível resultado da ação) na formulação do modelo e foram adicionados parâmetros, inspirados em modelos de ação dopaminérgica, que permitissem um desvio da estratégia ótima resultante da abordagem bayesiana. Os modelos construídos nessa estrutura foram simulados computacionalmente para diversos valores dos parâmetros, incluindo as memórias K e L do agente e da cadeia de Markov, respectivamente. Através de análises de correlação, esses resultados foram comparados aos dados experimentais, de um grupo de pesquisa do Instituto de Ciências Biomédicas da USP, referentes a tarefas de tomada de decisão envolvendo pessoas de diversas faixas etárias (de 3 a 73 anos) e cadeias de Markov de memórias 0, 1 e 2. Nessa comparação, concluiu-se que as diferenças entre grupos etários no experimento podem ser explicadas em nossa modelagem através da variação da memória K do agente crianças de até 5 anos mostram um limite K = 1, e as de até 12 anos mostram um limite K = 2 e da variação de um parâmetro de reforço de aprendizado dependendo do grupo e da situação de decisão à qual os indivíduos eram expostos, o valor ajustado desse parâmetro variou de 10% para baixo até 30% para cima do seu valor original de acordo com a abordagem bayesiana. / In this work we developed a simple probabilistic modeling framework that could describe the process of decision making in human agents that are presented with the task of predicting elements of a random sequence generated by a Markov chain with memory L. Such framework arised from a Bayesian approach in which the agent infers a probability distribution from a series of observations on the sequence and on its own answers, and considers that the agent\'s memory has length K. As a result of the Bayesian approach, the agent adopts an optimal strategy that consists in perseveration of the most likely alternative given the history of the last few trials; because of that and of experimental evidence that humans tend, in such kinds of problems, to adopt suboptimal strategies such as probability matching, variations on that model were developed in an attempt to have a closer description of the behavior adopted by humans. In that sense, the `shift\' (possible action taken by the agent on its response) and `reward\' (possible result of the action) variables were adopted in the formulation of the model, and parameters inspired by models of dopaminergic action were added to allow deviation from the optimal strategy that resulted from the Bayesian approach. The models developed in that framework were computationally simulated for many values of the parameters, including the agent\'s and the Markov chain\'s memory lengths K and L respectively. Through correlation analysis these results were compared to experimental data, from a research group from the Biomedical Science Institute at USP, regarding decision making tasks that involved people of various ages (3 to 73 years old) and Markov chains of orders 0, 1 and 2. In this comparison it was concluded that the differences between age groups in the experiment can be explained in our modeling through variation of the agent\'s memory length K children up to 5 years old exhibited a limitation of K = 1, and those up to 12 years old were limited to K = 2 and through variation of a learning reinforcement parameter depending on the group and the decision situation to which the candidates were exposed, the fitted value for that parameter ranged from 10% below to 30% above its original value according to the Bayesian approach.