教学课程资源库 更多>>

关键词： Computer science   Educational technology   Science education  

摘要： Programming skills are a vital part of many disciplines but can be challenging to teach and learn. Thus, the programming courses are considered difficult and a major stumbling block. To overcome these challenges, students could benefit from extensive individual support such as tutoring, but there are simply not enough qualified tutors available to meet rising demands. A potential solution is the development of intelligent tutoring systems (ITSs), which offer individualized, one-on-one instruction. Such systems can offer the support to make programming instruction more effective, scalable and reduce existing teachers' workloads. This dissertation demonstrates how conversational ITSs and the Socratic method of teaching can improve a novice's understanding of programming concepts and, in particular, the scaffolding of code comprehension processes. Furthermore, this work provides a novel method to automatically author a Socratic dialogue-based ITS. Indeed, two major outcomes of this work are a Socratic dialogue-based ITS and an automated dialogue authoring tool, which generates full Socratic dialogue from Java source code. The key objectives of this dissertation were, first, to determine whether the Socratic method would be effective at eliciting learners to engage in self-explanations with the help of the Socratic Tutor ITS and, second, to assess the quality of Socratic Author's auto-generated tutorial dialogue. Thus, the work presented here sought to answer two main research questions: (1) can a Socratic ITS lead to improved code comprehension? and (2) to what extent can Socratic dialogue be generated automatically? In sum, this research helps establish a relationship between code comprehension and the use of the Socratic method in learning computer programming. Furthermore, the work introduces a novel approach for generating Socratic dialogue from source code with examples for the Java programming language. The auto-authoring tool could help teachers and ITS develope

关键词： Computer science   Artificial intelligence  

摘要： Modern virtual reality (VR) head-mounted displays render more than 4 million pixels, and each pixel needs to be rasterized, lit, shaded, and colored. High-resolution displays and increased visual quality of VR content put a high burden on the graphics equipment; thus, only increasingly powerful graphics processing units (GPUs) handle the demand. Foveated rendering (FR) is a technique with the potential to reduce the required processing performance for graphics-intensive VR applications significantly. The technique adapts the rendering quality dynamically and only renders high-quality content at the user's focus point. Previous research proposed promising FR techniques. Research is often focused on the technical implementation of FR. However, it remains inconclusive whether or not it is beneficial in all situations. This research aims to investigate the performance gain break-even for a dynamic two-layered FR technique. The research approach is experimental. Three different scenes with changing geometrical or rendering complexity are analyzed on a low-end and a high-end GPU using rotating FR settings. The performance of the FR renderer and a standard renderer is profiled and compared to a theoretical model. The results indicate that a low-end mobile GPU benefits with FR implementation when the scene demands high occupancy with >121,000 vertices, ≥3 lights, and until a high-resolution circular region around the gaze point covers ≤30% of the scene. On the other hand, a high-end desktop GPU benefits with FR implementation when the scene is vast with >900,000 vertices, illuminated by ≥15 lights, and enabled with shadows while covering 10% of the scene with the high-resolution foveal region. The results meet the theoretical expectations and indicate that FR's performance gain is majorly affected by the number of threads through the graphics pipeline. In a nutshell, factors such as the scene characteristics, the order of complexity of the shading algorithms, graphics hardw

关键词： Computer Engineering   Computer science   Information Technology   Artificial intelligence  

摘要： Cloud computing has greatly impacted the scientific community and the end users. By leveraging cloud computing, small research institutions and undergraduate colleges are able to alleviate costs and achieve research goals without purchasing and maintaining all the hardware and software. In addition, cloud computing allows researchers to access resources as their teams require and allows real-time collaboration with team members across the globe. Nowadays however, users are easily overwhelmed by the wide range of cloud servers and instances. Due to differences between the cloud server platforms and between instances within the platform, users find it difficult to identify the right instance match for their application. Therefore, we propose the A2Cloud-Hierarchy (A2Cloud-H) framework that recommends Cloud instances to users for high-performance scientific computing. The framework comprises four components: training data collection, supervised learning (SL) module, unsupervised learning (USL) module, and a decision module. The training database comprise testing traces of previous application and Cloud instances; these are contributed by the scientific community. The SL module contains three popular supervised learning modules: logistic regression, support vector machine and random forest, which train using the database to qualitatively assess the instance performance for the target application. The USL module includes three collaborative filtering methods: application-based, instance-based and rank-based, which use the database to estimate the instances’ performance ratings for the target application. The decision module comprises multiple tiers of analytic hierarchy processing, which consolidate the instance recommendation from the SL and USL modules into a final instance recommendation. The model is trained and validated by 8 real-world applications on 20 Cloud instances, yielding more than 90% modeling accuracy. The recommendation and integration method proposed in

关键词： Computer science   Logic   Computer Engineering   Artificial intelligence  

摘要： Boolean Constraint Satisfaction Problems naturally arise in a variety of fields in Formal Methods and Artificial Intelligence. Constraint Solvers, the specialized software tools that solve them, are therefore a core enabling technology in industry and research. They are normally used as black-box components, applied to practical problems such as hardware and software verification, test generation, planning, synthesis and more. Based on classical algorithms that have been optimized over decades by researchers, there is a noticeable gap between Constraint Solvers and the technology of Deep Learning, which over the last decade found its way into countless domains, outperforming established domain-specific algorithms. This thesis aims to narrow this gap, and by using Deep Neural Networks, teach classical Constraint Solvers to "learn from experience''. The research I present in this thesis starts by addressing the challenge of representation, using a Graph Neural Networks based architecture to process propositional formulas as graphs. I will then show how to automatically learn a solver's branching heuristic by mapping it to a Markov Decision Process and training it using Deep Reinforcement Learning. I will present an implementation based on different two competitive solvers, and experiments showing automatically learned heuristics to outperform the state of the art in the two domains.

关键词： Electrical engineering   Computer science   Operations research  

摘要： The proposed research concerns the design of signal processing and communication techniques for sensor networks that focus on detection, estimation and parameter optimization. Specifically, we consider two problems, communication-efficient transmissions for detection and optimization in wireless sensor networks, and parameter estimation with unknown signals in radar sensor networks. Signal detection, estimation, and parameter optimization have been of great interest under the topic of sensor networking. Typically, spatially distributed sensor nodes carry their own limited energy sources. Hence, energy efficiency is an important consideration in sensor networking. In the first part of this thesis, we provide two approaches eliminating transmissions needed in a traditional optimal approach by using ordering and censoring. In ordering, the sensors with the most informative data transmit first and then transmissions stop when sufficient information is accumulated. We test the structure of a decomposable Gaussian Graphical Model (GGM) with this ordered transmission approach in a distributed scenario. We show that the ordered transmission approach requires fewer sensor transmissions and the savings can be large in cases of interest. Next, quickest change detection (QCD) with ordering in any decomposable graphical model is proposed to improve the communication efficiency without losing any detection performance compared to the optimal QCD algorithms. Different from ordering, in censoring, each node transmits only the informative observations but not necessarily in an ordered way. Based on censoring, we develop a new class of communication-efficient distributed learning algorithms called adaptive sparsified gradient descent (ASGD) methods with error correction that reduces the number of bits per communication as this is very important for large model parameter vector cases which are relevant for the popular deep learning approaches. A key feature of ASGD is that any given c

关键词： Computer engineering   Computer science   Artificial intelligence   Deep learning   Experiments   Brain research   Neural networks   Web browsers   Noise   Registration   Algorithms  

摘要： Electroencephalography (EEG) signals can be used for many purposes and has the potential to be adapted to various systems. When EEG is recorded from users, these studies are performed primarily in an indoor environment, while the user is stationary. This is due to the levels of noise that are experienced when recording EEG data, to minimize errors in the data. This thesis aims to adapt tasks that are performed indoors to an external environment by removing both noise and artefacts in EEG, using a 2D Convolutional Autoencoder (CAE). The data is recorded from subjects is passed into the 2D CAE to produce a reconstructed signal that removes the noise and artefacts. The experiment consists of an initial recording, where the subject sits stationary indoors for 60 seconds, to obtain a baseline. Afterwards, they perform movement-based tasks both indoors and outdoors for 60 seconds. These tasks consist of the following: stationary inside and outside with and without lifting a 10 lb dumbbell and movement inside and outside with and without lifting a 10lb dumbbell. The indoor movement recordings are used as the pure signals and the outdoor recordings are the noisy signals. Both are passed into the 2D CAE to produce a reconstructed signal that removes the noise from the outside recorded data. To verify results, a web application was developed to perform offline authentication. The reconstructed signal is matched against the data recorded while the user is moving in an internal environment, using the cosine similarity. If the two signals are closely related, the user will be able to authenticate into a web application. The matching score achieved were an average of 90% when performing the cosine similarity with the reconstructed signal to the inside data after preprocessing the data and 80% without preprocessing.

关键词： Online instruction   Probability   Deep learning   Computer science   Educational objectives   Curricula   Academic achievement   Distance learning   Neural networks   Skills   Artificial intelligence   Education   Mathematics  

摘要： Predicting student success has been a major task in student modeling. Specifically, in open-ended domains such as computer science classes, the student submissions contain more information, however they also require more advanced analysis methods to extract this information. Traditional student modeling approaches use knowledge components (KCs) to predict a student’s success on specific practiced skills. These approaches are useful and necessary in helping learning environments like Intelligent Tutoring Systems (ITS) personalize feedback, hints, and identify struggling students. However, when working with programming data, code features provide more information than skill tags representing KCs, and this information is not leveraged by traditional KC models. This work incorporates an implicit representation of KCs into a student model by including features extracted from students’ code with data from an undergraduate introductory programming course. This representation is then evaluated by using deep learning predictive models and investigated to see how well they are able to leverage code features to model student knowledge and compare and contrast against other learning models. The study shows a slight, but consistent improvement in models that use time-sequential data with even the simplest code features, implying that these aspects may improve student modelling.

摘要： Successful polices are usually implemented with the expectation of sustainment and education policies are no exception. In addition, education policies are often motivated by equity, even if the visions for equity are not explicitly described. The combination of ambiguously defined equity and the assumption of sustainability can lead to changes in, and dilution of, a policy’s original intents in ways that could leave a policy unrecognizable or ineffective in achieving its goals. This qualitative study sought to understand how equity is understood in the context of one state’s computer science (CS) policy, as well as expectations for sustaining the policy over time. The framework included examining equity using a distinction between inputs and outcomes, as well as a distinction between equal or differentiated inputs and outcomes. Conceptions of sustainability were explored by identifying what was sustained and how it was sustained. District CS leaders in one state were interviewed to understand how sustainment of the state’s CS policy was understood and how equity was conceptualized in the context of the policy. This study revealed that district leaders’ understanding of equity was dominated by beliefs that equalizing inputs would lead to differentiated, but proportional, student outcomes. Regarding sustainability, district leaders saw it as continuation of demand and supply of relevant CS courses in high schools. They described their role in sustainment as (1) increasing student enrollment, (2) fostering buy-in from families and school staff, and (3) recruiting, training, and retaining qualified teachers. The district leaders recognized that sustaining the policy required continual attention to it, particularly regarding buy-in and capacity in the form of qualified teachers and student interest. These findings highlight the value planning for sustainment by providing resources to foster buy-in and to maintaining teacher capacity, both with an emphasis on equity as t

关键词： Artificial intelligence   Logic   Computer science  

摘要： This thesis deals with the topic of modelling an agent’s beliefs about a dynamic world in a way that allows for changes in beliefs, including retracting of beliefs, based on the agent’s observations. We work within the field of knowledge representation, and represent the beliefs of the agent using a logical theory. In particular, we are concerned with representing what initial conditions the agent considers (im)plausible, what effects the agent thinks actions (im)plausibly have, and what processes in the environment the agent thinks have (im)plausibly occurred or will occur. Our approach uses the situation calculus, a standard knowledge representation framework for modelling action and change. Furthermore, we build on an existing framework in the situation calculus for modelling changing beliefs, where beliefs are determined using a plausibility ordering on situations. This supports modelling changing beliefs, since when the most plausible options are refuted by observations, the agent can fall back to the next most plausible options. Our concern is with how to specify this plausibility ordering using a logical theory. We propose to define the ordering by counting certain properties of situations, indicated by distinguished predicates, which we call “abnormality” predicates. This is inspired by how minimization of abnormalities has been used in circumscription, an approach to default reasoning. We show how beliefs about plausible and implausible action effects can be represented by having the axioms describing effects refer to abnormalities. Furthermore, we extend the account of belief to allow for beliefs about ongoing exogenous processes, described by a program (written in ConGolog, a standard programming language for use with the situation calculus). We show how having these programs refer to abnormalities allows for representing plausible and implausible environment behavior. Finally, we present a formal definition of “knowing how” to achieve goals, in terms of

关键词： Transportation   Electrical engineering   Computer science   Artificial intelligence  

摘要： The Association for Unmanned Vehicle Systems International (AUVSI) estimates that, by 2025, more than 100,000 jobs will be created with an economic impact of $82 billion in the commercial drone market in the US. In the field of Transportation, drones have proven useful, especially for collecting video data through mounted cameras. Unfortunately, further processing of such videos to extract relevant transportation metrics can be very challenging as traditional data extraction methodologies rely on background subtraction which performs best when the camera is fixed. Unlike Background Subtraction, Machine Learning object detectors like You Only Look Once (YOLO) do not depend on motion to detect objects of interest in images; however, they suffer from high misdetections due to varying sizes of objects. This framework is an autonomous system for extracting temporal vehicle track information from unstabilized videos combining computer vision methodologies and machine learning. The extracted information is provided in a quarriable format for further use in traffic analysis.