Made available in DSpace on 2014-06-12T18:08:23Z (GMT). No. of bitstreams: 2 arquivo9473_1.pdf: 920450 bytes, checksum: 8de41a22d93f1f66a4f3481b45626f98 (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2012 / Devido a maior oferta de petróleos pesados e alto grau de contaminantes que os derivados deste possuem, os processos de hidrorrefino têm recebido atenção especial ao longo dos últimos por possibilitar a remoção de contaminantes e melhorar a margem de lucro das refinarias por tonar possível a obtenção de derivados de maior valor agregado. Entre esses o processo de hidrotratamento (HDT), no qual ocorre uma série de reações que utilizam o gás hidrogênio como reagente, foi o foco de estudo deste trabalho. Ao ser aplicado em correntes de Diesel o HDT realiza a remoção de contaminantes como enxofre e nitrogênio, aumentando a qualidade do mesmo. A unidade de HDT tem como principal equipamento o reator, que consiste em um leito com partículas sólidas, onde gás e líquido escoam em fluxo co-corrente ou em contracorrente. Apesar deste processo já ser maduro, o crescente aumento nas exigências de mercado demandam por melhorias no mesmo, a fim de atingir uma rentabilidade cada vez maior. Desta forma o uso de inferenciadores na estimação das variáveis tornaria possível o melhor acompanhamento do processo como também a implementação de novas estratégias de controle. Visto a relevância desse tema o presente trabalho abordou o desenvolvimento de observadores de estado para o reator do processo de HDT, para isto foi necessário a aquisição de dados do processo, o que foi conseguido através de um modelo matemático do reator, o qual foi denominado como planta virtual. Esta forneceu os dados para treinamento e validação dos inferenciadores aqui estudados: as redes neuronais e a neuro-fuzzy. No decorrer do trabalho foi definido o tempo de amostragem e o período de excitação do sinal através da menor constante de tempo. Para treinamento dos inferenciadores foi utilizado dois bancos de dados distintos, um com tempo de amostragem de 50s, onde este foi obtido pelo método da constante de tempo, e outro com amostragem de 10 minutos, em que as seguintes variáveis foram inferenciadas: concentração de compostos sulfurados, nitrogenados e olefinas na saída do reator. Dessas o melhor resultado foi obtido na inferência da concentração de compostos sulfurados realizada através da Rede Neuronal. Foi escolhida esta rede neuronal na implementação de um controlador PID e como modelo interno de um controlador NNMPC. O controlador PID cuja variável de controle foi à concentração de sulfurados foi chamado de controlador PID inferencial e os resultados deste se mostraram melhores do que os resultados obtidos pelo controlador NNMPC

Robots should offer a human-like level of dexterity when handling objects if humans are to be replaced in dangerous and uncertain working environments. This level of dexterity for human-like manipulation must come from both the hardware, and the control. Exact replication of human-like degrees of freedom in mobility for anthropomorphic robotic hands are seen in bulky, costly, fully actuated solutions, while machine learning to apply some level of human-like dexterity in underacted solutions is unable to be applied to a various array of objects. This thesis presents experimental and theoretical contributions of a novel neuro-fuzzy control method for dextrous human grasping based on grasp synergies using a Human Computer Interface glove and upgraded haptic-enabled anthropomorphic Ring Ada dexterous robotic hand. Experimental results proved the efficiency of the proposed Adaptive Neuro-Fuzzy Inference Systems to grasp objects with high levels of accuracy.

The focus of this thesis was to measure the intracellular voltage of a living neural cell using a silicon transistor. The coupling of neurological tissues with silicon devices is of interest to the fields of neurology, neuroscience, electrophysiology and cellular biology. In previous work by Peter Fromherz, single neurons were successfully coupled to transistors [1]. This thesis aims to show proof of concept of the fabrication of a simple neuro-silicon interface using wafer processing methods currently available at Cal Poly. The types of transistors and cells used, the methods for dissecting and preparing the cells, the electrophysiology methods for validating the experiments, and portions of the design of the junction were based on Fromherz’s 1991 work. Other aspects were revised to be compatible with technologies available at Cal Poly. Leech Retzius cells were isolated and cultured from Hirudo Medicinalis and joined to the gate oxide of a P-channel field effect transistor using SU-8 photoresist wells treated with poly-l-lysine. Transistors were operated in strong inversion and source-drain currentfluctuations were observed that correlated with action potentials of the current clamped Retzius cell. Further work is needed to develop better junctions that can reliably couple action potentials. This work lays a foundation for neuro-silicon interface fabrication at Cal Poly.

No

This dissertation investigates the turn of the century fascination with synesthesia and efforts by Modernist artists and writers to produce cross-modal works that attempt to defy sensory boundaries. Works of impossible art are artistic and literary experiments with style and form that develop out of the realism and naturalism of the nineteenth century, to be sure; they are also conceived of by their creators as scientific experiments that test what is possible at the limits of perception. Accordingly, while my work is situated within the field of aesthetics, I take a neuroscientific approach to aid in understanding the modes of perception these works are attempting to explore. My project applies the findings of recent neuroscientific studies into clinical synesthesia as a guide for thinking about these Modernist works. The methodology of neuro-aesthetics allows me to develop a theory of sensory mimesis. Sensory mimesis is a holistic approach to explaining phenomenological experience that depends on a sensory semantics, more fundamental and more comprehensive than a linguistic semantics, that I propose filters our access to the world. What we ultimately learn from impossible art is that the range of neurodiversity in humans is broader than we tend acknowledge or appreciate. The notoriously indefinable and uncategorizable character of queer theory is an applicable framework to match the innumerable neurocognitive possibilities that are actually available. To this end, my dissertation suggests that a shift to a neuro-queer-aesthetic paradigm would not only expand human perceptive possibilities, but also enable compassionate engagement within and among our diverse communities. / English

No description available.

Titre de l'écran-titre (visionné le 29 novembre 2023) / L'apprentissage automatique a considérablement progressé dans le domaine de la recherche en neurosciences, mais son application pose des défis en raison des différences entre les principes biologiques du cerveau et les méthodes traditionnelles d'apprentissage automatique. Dans ce contexte, le projet présenté propose NeuroTorch, un pipeline convivial d'apprentissage automatique spécialement conçu pour les neuroscientifiques, afin de relever ces défis. Les objectifs clés de ce projet sont de fournir une librairie d'apprentissage profond adaptée aux neurosciences computationnelles, d'implémenter l'algorithme eligibility trace forward propagation (e-prop) pour sa plausibilité biologique, de comparer les réseaux de neurones continus et à impulsions en termes de résilience, et d'intégrer un pipeline d'apprentissage par renforcement. Le projet se divise en plusieurs parties. Tout d'abord, la théorie des dynamiques neuronales, des algorithmes d'optimisation et des fonctions de transformation d'espaces sera développée. Ensuite, l'attention sera portée sur la conception du pipeline NeuroTorch, incluant l'implémentation de l'algorithme e-prop. Les résultats de la prédiction de séries temporelles d'activité neuronale chez le poisson-zèbre seront présentés, ainsi que des observations sur la résilience à l'ablation des réseaux obtenus. Enfin, une section sera consacrée à l'exploration du pipeline d'apprentissage par renforcement de NeuroTorch et à la validation de son architecture dans l'environnement LunarLander de Gym. En résumé, les modèles à impulsions de NeuroTorch ont atteint des précisions de 96,37%, 85,58% et 74,16% respectivement sur les ensembles de validation MNIST, Fashion-MNIST et Heidelberg. De plus, les dynamiques leaky-integrate-and-fire with explicit synaptic current - low pass filter (SpyLIF-LPF) et Wilson-Cowan ont été entraînées avec succès à l'aide de l'algorithme e-prop sur des données neuronales expérimentales du ventral habenula du poisson-zèbre, obtenant respectivement des valeurs de pVar de 0,97 et 0,96. Les résultats concernant la résilience indiquent que l'application de la loi de Dale améliore la robustesse des modèles en termes d'ablation hiérarchique. Enfin, grâce au pipeline d'apprentissage par renforcement de NeuroTorch, différents types d'agents inspirés des neurosciences ont atteint le critère de réussite dans l'environnement LunarLander de Gym. Ces résultats soulignent la pertinence et l'efficacité de NeuroTorch pour les applications en neurosciences computationnelles. / Machine learning has made significant advancements in neuroscience research, but its application presents challenges due to the differences between the biological principles of the brain and traditional machine learning methods. In this context, the presented project proposes NeuroTorch, a comprehensive machine learning pipeline specifically designed for neuroscientists to address these challenges. The key objectives of this project are to provide a deep learning library tailored to computational neuroscience, implement the eligibility trace forward propagation (e-prop) algorithm for biological plausibility, compare continuous and spiking neural networks in terms of resilience, and integrate a reinforcement learning pipeline. The project is divided into several parts. Firstly, the theory of neural dynamics, optimization algorithms, and space transformation functions will be developed. Next focus will be on the design of the NeuroTorch pipeline, including the implementation of the e-prop algorithm. Results of predicting a time series of neuronal activity in zebrafish will be presented, along with observations on the resilience to network ablations obtained. Finally, a section will be dedicated to exploring the NeuroTorch reinforcement learning pipeline and validating its architecture in the LunarLander environment of Gym. In summary, NeuroTorch spiking models achieved accuracies of 96.37%, 85.58%, and 74.16% on the MNIST, Fashion-MNIST, and Heidelberg validation sets, respectively. Furthermore, the leaky-integrate-and-fire with explicit synaptic current - low pass filter (SpyLIF-LPF) and Wilson-Cowan dynamics were successfully trained using the e-prop algorithm on experimental neuronal data from the ventral habenula of zebrafish, achieving pVar values of 0.97 and 0.96, respectively. Results regarding resilience indicate that the application of the Dale law improves the robustness of models in terms of hierarchical ablation. Lastly, through the NeuroTorch reinforcement learning pipeline, different types of neuroscience-inspired agents successfully met the success criterion in the Gym's LunarLander environment. These results highlight the relevance and effectiveness of NeuroTorch for applications in computational neuroscience.

In the realm of natural language programming, translating free-form sentences in natural language into a functional, machine-executable program remains difficult due to the following 4 challenges. First, the inherent ambiguity of natural languages. Second, the high-level verbose nature in user descriptions. Third, the complexity in the sentences and Fourth, the invalid or semantically unclear sentences. Our first solution is a Large Language Model (LLM) based Artificial Intelligence driven assistant to process free-form sentences and decompose them into sequences of simplified, unambiguous sentences that abide by a set of rules, thereby stripping away the complexities embedded within the original sentences. These resulting sentences are then used to generate the code. We applied the proposed approach to a set of free-form sentences written by middle-school students for describing the logic behind video games. More than 60% of the free-form sentences containing these problems were sufficiently converted to sequences of simple unambiguous object-oriented sentences by our approach. Next, the thesis also presents "IntentGuide," a neuro-symbolic integration framework to enhance the clarity and executability of human intentions expressed in freeform sentences. IntentGuide effectively integrates the rule-based error detection capabilities of symbolic AI with the powerful adaptive learning abilities of Large Language Model to convert ambiguous or complex sentences into clear, machine-understandable instructions. The empirical evaluation of IntentGuide performed on natural language sentences written by middle school students for designing video games, reveals a significant improvement in error correction and code generation abilities compared to previous approach, attaining an accuracy rate of 90%. / Master of Science / Imagine if you could talk to machines in everyday language and they could understand exactly what you meant, turning your words into programs that do exactly what you describe. That's the goal of the thesis. We've developed a system that helps machines make sense of the kind of free-form language that people, especially students, use when they describe what they want a computer to do. Understanding and converting everyday language into computer code is a complex challenge, primarily because the way we naturally speak can be vague, overly detailed, or just complex. This thesis presents a new tool using artificial intelligence that helps break down and simplify these sentences. By transforming them into clearer, rulefollowing instructions, this tool makes it easier for machines to understand and execute the tasks we describe. The technology was tested using descriptions from middle-school students on how video games should work. Over 60% of these complex or unclear descriptions were sufficiently converted into straightforward instructions that a machine could use. Additionally, a new system called "IntentGuide" was introduced, combining traditional AI methods with advanced language models to improve how effectively machines can interpret and act on human instructions. This improved system showed a 90% accuracy in understanding and correcting errors in the students' game descriptions, marking a significant step forward in helping computers better understand us.

In the past decade, reinforcement learning (RL) has achieved breakthroughs across various domains, from surpassing human performance in strategy games to enhancing the training of large language models (LLMs) with human feedback. However, RL has yet to gain widespread adoption in mission-critical fields such as healthcare and autonomous vehicles. This is primarily attributed to the inherent lack of trust, explainability, and generalizability of neural networks in deep reinforcement learning (DRL) agents. While neural DRL agents leverage the power of neural networks to solve specific tasks robustly and efficiently, this often comes at the cost of explainability and generalizability. In contrast, pure symbolic agents maintain explainability and trust but often underperform in high-dimensional data. In this work, we developed a method to distill explainable and trustworthy agents using neuro-symbolic AI. Neuro-symbolic distillation combines the strengths of symbolic reasoning and neural networks, creating a hybrid framework that leverages the structured knowledge representation of symbolic systems alongside the learning capabilities of neural networks. The key steps of neuro-symbolic distillation involve training traditional DRL agents, followed by extracting, selecting, and distilling their learned policies into symbolic forms using symbolic regression and tree-based models. These symbolic representations are then employed instead of the neural agents to make interpretable decisions with comparable accuracy. The approach is validated through experiments on Lunar Lander and Pong, demonstrating that symbolic representations can effectively replace neural agents while enhancing transparency and trustworthiness. Our findings suggest that this approach mitigates the black-box nature of neural networks, providing a pathway toward more transparent and trustworthy AI systems. The implications of this research are significant for fields requiring both high performance and explainability, such as autonomous systems, healthcare, and financial modeling.

Titre de l'écran-titre (visionné le 31 octobre 2023) / L'imagerie par résonance magnétique fonctionnelle (IRMf) est une technique d'imagerie non invasive. Celle-ci utilise les champs magnétiques afin de mesurer les changements hémodynamiques induits par le mécanisme du couplage neurovasculaire dans une région du cerveau donnant une mesure indirecte de l'activité neuronale. La réalisation d'études IRMf sur des souris éveillées pose diverses difficultés et défis techniques nécessitant l'utilisation de sédatifs ou d'anesthésiques. Or, certains facteurs comme le stress et l'état d'anesthésie sont reconnus pour altérer la fonction cérébrale affectant simultanément la fidélité des résultats recueillis. Le but de ce projet est ainsi d'étudier la faisabilité de l'imagerie par résonance magnétique fonctionnelle à 1 tesla chez la souris éveillée. D'abord, une méthode de fixation a été développée afin de restreindre les mouvements d'une souris permettant de l'imagerie anatomique et fonctionnelle in vivo. Un paradigme en bloc alternant une période de repos et une période de stimulation a été élaboré en utilisant un mélange gazeux de dioxyde de carbone et d'oxygène à titre de stimulation. Le tout, afin d'induire des changements sanguins comparables à ceux provoqués par le couplage neurovasculaire. Ensuite, une analyse statistique, sur les images fonctionnelles, a permis d'obtenir deux cartes d'activation en comparant les deux différents blocs avec stimulation et au repos. Finalement, les résultats obtenus à l'IRMf sont comparés avec la technique d'imagerie optique intrinsèque pour vérifier la concordance des réponses mesurées par diverses méthodes d'imagerie. Ainsi, l'obtention de ces cartes montre qu'une étude IRMf sur une IRM 1T avec souris éveillée est possible. / Functional magnetic resonance imaging (fMRI) is a non-invasive imaging technique that uses magnetic fields to measure hemodynamic changes induced by the mechanism of neurovascular coupling in a region of the brain giving an indirect measure of neuronal activity. Performing fMRI studies on awake mice poses various technical difficulties and challenges requiring the use of sedatives or anesthetics. However, certain factors such as stress and the state of anesthesia are known to alter brain function, simultaneously affecting the fidelity of the collected results. The aim of this project is to study the feasibility of functional magnetic resonance imaging at 1 tesla in awake mice. First, a mouse holder was developed to restrict the movements of a mouse's head allowing for anatomical and functional imaging in vivo. A block paradigm of alternating a period of rest and a period of stimulation was developed using a gaseous mixture of carbon dioxide and oxygen as stimulation in order to induce blood changes comparable to those caused by neurovascular coupling. Then, a statistical analysis on those functionals images made it possible to obtain two activation maps by comparing the two different blocks of stimulation and at rest. Finally, the results obtained by fMRI are compared with the intrinsic optical imaging technique to verify the concordance of the responses through different imaging methods. Thus, obtaining these maps shows that an fMRI study on a 1T MRI with awake mice is possible.