教学课程资源库 更多>>

关键词： Computer science   Artificial intelligence   Morphology   Neurosciences   Operations research   Web Studies  

摘要： How can we make an agent that thinks like us humans? An agent that can have proprioception, intrinsic motivation, identify deception, use small amounts of energy, transfer knowledge between tasks and evolve? This is the problem that this thesis is focusing on. Being able to create a piece of software that can perform tasks like a human being, is a goal that, if achieved, will allow us to extend our own capabilities to a very high level, and have more tasks performed in a predictable fashion. This is one of the motivations for this thesis. To address this problem, we have proposed a modular architecture for Reinforcement Learning computation and developed an implementation to have this architecture exercised. This software, that we call mHBP, is created in Python using Webots as an environment for the agent, and Neo4J, a graph database, as memory. mHBP takes the sensory data or other inputs, and produces, based on the body parts / tools that the agent has available, an output consisting of actions to perform. This thesis involves experimental design with several iterations, exploring a theoretical approach to RL based on graph databases. We conclude, with our work in this thesis, that it is possible to represent episodic data in a graph, and is also possible to interconnect Webots, Python and Neo4J to support a stable architecture for Reinforcement Learning. In this work we also find a way to search for policies using the Neo4J querying language: Cypher. Another key conclusion of this work is that state representation needs to have further research to find a state definition that enables policy search to produce more useful policies. The article “REINFORCEMENT LEARNING: A LITERATURE REVIEW (2020)” at Research Gate with doi 10.13140/RG.2.2.30323.76327 is an outcome of this thesis.

关键词： Engineering   Computer science  

关键词： Robotics   Computer science   Environmental health   Optics   Electrical engineering  

摘要： Floating trash objects are very commonly seen on water bodies such as lakes, canals and rivers. With the increase of plastic goods and human activities near the water bodies, these trash objects can pile up and cause great harm to the surrounding environment. Using human workers to clear out these trash is a hazardous and time-consuming task. Employing autonomous robots for these tasks is a better approach since it is more efficient and faster than humans. However, for a robot to clean the trash objects, a good detection algorithm is required. Real-time object detection on water surfaces is a challenging issue due to nature of the environment and the volatility of the water surface. In addition to this, running an object detection algorithm on an on-board processor of a robot limits the amount of CPU consumption that the algorithm can utilize. In this thesis, a computationally low cost object detection approach for robust detection of trash objects that was run on an on-board processor of a multirotor is presented. To account for specular reflections on the water surface, we use a polarization filter and integrate a specularity removal algorithm on our approach as well. The challenges faced during testing and the means taken to eliminate those challenges are also discussed. The algorithm was compared with two other object detectors using 4 different metrics. The testing was carried out using videos of 5 different objects collected at different illumination conditions over a lake using a multirotor. The results indicate that our algorithm is much suitable to be employed in real-time since it had the highest processing speed of 21 FPS, the lowest CPU consumption of 37.5% and considerably high precision and recall values in detecting the object.

关键词： Mechanical engineering   Computer science   Computer Engineering   Thermodynamics  

摘要： Uneven temperature distribution poses a big challenge to reflow process of surface mount components. Particularly, when components with different thermal mass are soldered on a large size PCB, temperature gradient across the board can cause various soldering defects. The formation of a robust solder joint is closely associated with its temperature profile which is hard to be measured experimentally. Thus, numerical simulation is an effective method to predict the solder joint temperature which can be used as a reference to optimize the reflow profile. In this study, CFD modeling was developed to obtain solder temperature profile in reflow process. Linear and soak profiles were both investigated and compared. The optimized thermal profiles and corresponding temperature setting of reflow oven can contribute to actual experiment and production. Besides, an innovative method was proposed to decrease solder temperature differences among components.

关键词： Computer science   Biomedical engineering   Artificial intelligence  

摘要： In an ordinary system, doctors use an acoustic stethoscope to detect any aberration in a heartbeat and predict atypical conditions of the human heart. One major problem is that the frequency range and intensity of the heart sounds are flat, and the sound may contain noise. Hence, even a cardiac specialist doctor may encounter difficulties in analyzing the heart sound perfectly. This paper describes a new methodology to detect heart diseases by examining heart sounds in real-time. Assuming the “gut” sounds to be input data, the methodology described here uses a deep learning approach to figure out whether a patient has any disease or is healthy. It is aimed to design a real-time system methodology integrated with an electronic stethoscope and a software solution known as a heartbeat audio classifier. Our proposed system solution should differentiate normal heartbeats and heart murmurs with a prediction of probable heart problem type in real-time. We believe our approach assists in reducing the cardiac arrest rate.

关键词： Computer science   Artificial intelligence  

摘要： Building agents with general skills that can be applied in a wide range of settings has been a long-standing problem in machine learning. The most popular framework for training agents to make sequential decisions in order to maximize reward in a given environment is Reinforcement Learning (RL). Over the last decade, deep reinforcement learning, where RL agents are parameterized by neural networks, has achieved impressive results on a number of tasks, from games such as Atari [Mnih et al. 2013], Go [Silver et al. 2016], StarCraft [Vinyals et al. 2019], or Dota [Berner et al. 2019], to continuous control tasks with applications in robotics [Lillicrap et al. 2016, Gu et al. 2017]. However, current RL agents are prone to overfitting and struggle to generalize when even minor perturbations are applied to the training environment [Rajeswaran et al. 2017, Zhang et al. 2018a, Zhang et al. 2018b]. This hinders progress on real-world applications such as autonomous vehicles or home robots, where agents need to deal with a large variety of scenarios. In this thesis, we introduce several methods for improving the versatility of deep reinforcement learning agents. We start by studying the problem of zero-shot generalization to new instances of a task after training on a limited number of environments. In this setting, agents are trained directly from partial observations of the environment in the form of images. We first propose an approach for regularizing the policy and value function of a RL agent and automatically finding an effective type of data augmentation for a given task. In addition, we demonstrate that a naïve application of data augmentation in RL can hurt performance in practice and is not theoretically sound for a certain class of algorithms. We also identify that there is an asymmetry between the information needed to represent the optimal policy and the true value function, which leads to overfitting when using standard deep RL algorithms. As a step towards sol

关键词： Bioinformatics   Computer science   Microbiology   Information technology   Datasets   Phylogenetics   Experiments   Databases   Taxonomy   Genomes   Algorithms   Genes  

摘要： Complex microbial communities play a crucial role in environmental and human health. Traditionally, microbes have been studied by isolating and culturing them, missing organisms that cannot grow in standard laboratory conditions, and losing information about microbe-microbe interactions. With affordable high- throughput sequencing, a new field called metagenomics has emerged, that studies the genomic content of the microbial community as a whole. Metagenomics enables researchers to characterize complex microbial communities, however, many computational challenges remain with downstream analyses of large sequencing datasets. Here, we explore some fundamental problems in metagenomics and present simple algorithms and open-source software tools that implement these solutions. In the first section, we focus on using a reference database of known organisms (and genomic segments within) to taxonomically classify sequences and estimate abundances of taxa in a metagenomic sample. We developed a “BLAST outlier detection” algorithm that identifies significant outliers within database search results. We extended this method and developed ATLAS, which uses significant database hits to group sequences in the database into partitions. These partitions capture the extent of ambiguity within the classification of a sample. Besides taxonomically classifying sequences, we also explored the problem of taxonomic abundance profiling, i.e., estimating the abundance of different species in the community. We describe TIPP2, a marker gene-based abundance profiling method, which combines phylogenetic placement with statistical techniques to control classification accuracy. TIPP2 includes an updated set of reference packages and several algorithmic improvements over the original TIPP method. Next, we explore how to reconstruct genomes from metagenomic samples. Despite advances in metagenome assembly algorithms, assembling reads into complete genomes is still a computationally challenging prob

关键词： Medicine   Biology   Bioinformatics   Artificial intelligence   Biochemistry   Chemistry   Computer science   Genetics   Morphology  

摘要： Third-generation long-read sequencing has enhanced the characterization of genomic structural variations (SV) in humans. However, due to its high cost and low throughput, its utilization in large-scale genetic studies or routine clinical investigations remains challenging. This thesis presents an innovative bioinformatics approach to overcome these limitations to allow cost-efficient and comprehensive discovery of SVs in the human genome. I develop NanoVar, an open-source bioinformatics software that enables the accurate discovery of SV at a lower sequencing depth and cost. At a sequencing depth of 8X, NanoVar outperformed existing tools in variant calling accuracy in simulated (F1 > 0.92) and real benchmark datasets. By employing NanoVar in a cohort of primary acute myeloid leukemia (AML) (n=17) and normal samples (n=7), my pilot study revealed 80 promising AML-associated variants. I hope that NanoVar may advance clinical SV research and expand our search for more druggable variants for precision medicine.

关键词： Computer science   Computer engineering   Artificial intelligence  

关键词： Computer science   Artificial intelligence  

摘要： Spatial augmented reality (SAR) aims to superimpose interactive visual contents onto the real world using projector-camera systems (ProCams). Compared with traditional augmented reality techniques, it has the potential to share virtual contents across multiple users without requiring additional displays, such as smartphones and augmented reality glasses. Although abundant efforts have been dedicated to SAR applications, how to create precise and satisfactory immersive and natural effects remains challenging. Moreover, SAR privacy and security have gained increasing attention as the advancement of deep learning-based visual recognition approaches. In this dissertation, we propose techniques of visualization and security in SAR, and in particular, we focus on key challenges such as precise geometric calibrations, projector compensation, image-based relighting, shape reconstruction, physical mail privacy and projector-based adversarial attacks. We first develop a single-shot-per-pose structured light-based ProCams calibration method for imperfect planar calibration boards. Then, we propose end-to-end projector compensation methods CompenNet, CompenNet++ and CompenNeSt++ to cancel photometric and geometric distortions. Afterwards, we present a unified approach named DeProCams for simultaneous relighting, compensation and shape reconstruction. Finally, focusing on the privacy and security of SAR applications, we propose Neural-STE and SPAA for privacy attacks on physical mail and projector-based adversarial attacks, respectively.