21793 research outputs found
Sort by
On Using Simulated Annealing in Training Deep Neural Networks
Get PDFIn deep learning, overfitting is a major problem that makes it difficult for a neural network to perform well on new data. This issue is especially prevalent in low-data regimes, or when training for too many epochs. Iterative learning methods have been devised to improve the generalization performance of neural networks when trained for a prolonged duration. These techniques periodically reduce the training accuracy of a network which is called forgetting. The primary objective of the forgetting stage is to allow the network to learn more from the same data and surpass its previous performance over the long run.
In this thesis, we propose a new forgetting technique motivated by simulated annealing. Although simulated annealing is a powerful tool in optimization, its application in deep learning has been overlooked. In our study, we highlight the potential of this method in deep learning and illustrate its usefulness through experiments. Essentially, we select a subset of layers to undergo brief periods of gradient ascent, followed by gradient descent. In the first scenario, we utilize Simulated Annealing in Early Layers (SEAL) during the training process. Through extensive experiments on the Tiny-ImageNet dataset, we demonstrate that our method has a much better prediction depth, in-distribution, and transfer learning performance compared to the state-of-the-art works in iterative training. In the second scenario, we expand the application of simulated annealing beyond the realms of classification and computer vision, by employing it in text-to-3D generative methods. In this scenario, we apply simulated annealing to the entire network and illustrate its effectiveness compared to normal training. These two scenarios collectively demonstrate the potential of simulated annealing as a valuable tool for optimizing deep neural networks and emphasize the need for further exploration of this technique in the literature
clip-mesh: generating textured meshes from text using pretrained image-text models
Get PDFThe following thesis introduces a novel technique for generating textured mesh models without any 3D supervision based solely on a text prompt. This is done by deforming the control shape of a limit subdivided surface along with its texture and normal map to match an input text prompt. The generated mesh asset can be easily integrated into games or modeling applications that rely on widespread rasterization based rendering techniques. The approach relies on a pre-trained Contrastive Language-Image Pre-Training (CLIP) model to compare the input text prompt with differentiably rendered images of our initialized 3D model. Unlike previous works that focused on stylization or required training of generative models, it performs optimization on mesh parameters directly to generate shape, texture, or both. To ensure that the optimization produces plausible meshes and textures, this work introduces several techniques including image augmentations, camera tuning and use of a pre-trained prior that generates CLIP image embeddings given a text embedding. Overall, this method offers a promising solution for zero-shot generation of 3D models, demonstrating the potential of CLIP-based techniques for the field of computer graphic
"religion is one thing, politics another:" examining the religiosities of cuban immigrants in montreal
Get PDFThis thesis examines the impact of Fidel Castro’s regime on the religious lives of Cuban immigrants currently residing in Montreal. Outlining a history of the socio-political milieu which allowed Castro to come to power, as well as the subsequent regime he instituted for nearly sixty years, this project focuses on the impact of Castro’s restrictive policies on religious practice throughout the island. By conducting life story interviews with ten Cuban immigrants who now call Montreal home, this paper argues that the Revolutionary government’s oppression of religious practice throughout the island produced religious illiteracy among the population due to a severe lack of religious education coupled with laws forbidding religious practice. To define religious illiteracy among the interlocutors, I move away from discussing a general lack of religious knowledge, as some do have knowledge about religion to a certain degree, especially in regards to their own religion. Instead, I am referring to a lack of comfort and a lack of breadth of knowledge which, in turn, manifests as intolerance against certain religions, continuing the anti-religious stance of the Cuban Revolutionary government, regardless of the interlocutors’ political views towards the State. Engaging with scholars who have focused on the intersection between migration stories and oral histories, as well as scholars whose works problematize or explore the religious diversity of Montreal, I aim to bring the religious plurality of Montreal and Cuba in conversation with the comparative experiences that emerge from the interlocutors’ respective migrations to Canada. Also working with scholars who have focused on the tempestuous relationship between religion (the Catholic Church, Protestant Christianity, and Afro-Cuban religions) and the Revolutionary government, this paper concludes by considering the interlocutors’ perspectives of Montreal as a religiously diverse city as well as the impact of living in such a space on their understandings of religion(s), their personal religious practices, and beliefs
Framework for Multi-Purpose Utility Tunnel Location Selection Considering Social Costs
Get PDFThe unsustainable way of burying utilities has resulted in difficulties in regular maintenance, which is one of the main reasons for the poor state of these utilities. Accessing these utilities through open-cut excavation leads to detrimental socioeconomic and environmental impacts, which can be quantified as social costs incurred during the asset's lifecycle. To reduce social costs, synchronized interventions of collocated assets (e.g., water and sewer pipes and pavements) can minimize excavations in a specific street segment over time for preventive maintenance. Few studies have prioritized street segments based on the socioeconomic impacts of intervention activities, such as street closures. Additionally, synchronized interventions are not a long-term solution as they do not eliminate the need for future excavations.
A more sustainable solution is the Multi-Purpose Utility Tunnels (MUTs), as they integrate all underground utilities in one accessible tunnel. MUTs eliminate the need for future excavations and their associated costs, as well as the resulting socioeconomic impacts. Meanwhile, both the synchronized interventions and the MUT are not generally practiced. Although several MUTs have been implemented in different parts of the world, their locations have either been politically influenced or selected to preserve heritage sites. In some cases, MUTs are built to take advantage of some opportunities, such as a newly developed area. Nevertheless, choosing the street segments for MUTs is affected by several criteria with different spatial characteristics. Combining these characteristics and managing their trade-offs determine the ranking of alternative MUT locations. Therefore, a systematic approach for MUT location selection that is based on the spatial characteristics of the criteria, as well as the lifecycle costs of each alternative is needed. Furthermore, the social cost of maintaining, repairing, or replacing utility assets during synchronized interventions or MUT implementation can be reduced by predicting the closure of street segments based on the need for intervention determined by asset conditions.
The main goal of this research is to develop a framework for MUT location selection considering social costs. To achieve this goal, the specific objectives are: (1) Establishing the relationships between intervention activities and their socioeconomic impacts; (2) Classifying the conditions of different spatially collocated underground municipal assets (i.e., pavements, water and sewer pipes) within a segment; (3) Determining street closures based on the synchronized or unsynchronized interventions at the segment level; (4) Developing a multi-criteria decision-making model (MCDM) for MUT location selection; and (5) Optimizing the location selection of MUTs considering agency and social lifecycle costs.
The main contributions developed in the context of this thesis are: (1) Developing a geospatial visual analytics model that supports the understanding of the socioeconomic impacts of unsynchronized intervention practices; (2) Developing an ML method for systematic condition classification of different spatially collocated underground municipal assets (i.e., pavements, water and sewer pipes) within a segment. To the best of our knowledge, there is no existing research about determining street closures based on the combined conditions of spatially collocated municipal infrastructure assets at the segment level; (3) Applying a heuristic approach for determining street closures based on the synchronized or unsynchronized interventions at the segment level induced by combining the interventions of individual assets within each segment; (4) Defining a comprehensive MCDM model that identifies and quantifies the criteria that influence the MUT location selection using subjective and objective methods; (5) Defining a multi-objective optimization model that was able to identify the potential MUT locations and a multi-year plan for MUT implementation that offers lifecycle savings; (6) Developing a systematic method for comparing the results of the MUT optimization with those of the synchronized interventions. This comparison is based on the agency and social LCCs, and the network deterioration generated by the MUTs and synchronized method of utility interventions at the network and segment levels; and (7) Developing three regression models for capturing the social cost of both alternatives at the network and segment levels
Vibration-based wheel-terrain slip detection for skid-steer rovers
Get PDFTo fulfill their mission properly, planetary exploration rovers must often be able to travel long distances and traverse various terrain types. Some terrain types and topologies may present traversability challenges. In difficult situations, such as slopes and loose soil, rover wheel slip may increase to a level leading to entrapment risks. Autonomous slip detection allows a rover to detect potentially dangerous terrain and take precautions. While visual odometry slip estimation solutions exist, numerous external factors, such as luminance, haze and shadows, may negatively impact quality of imaging sensor data and consequently slip estimation. Visual odometry also requires significant computational resources. Previous studies have shown promise in the use of Machine Learning algorithms to process IMU-measured vibration data to detect and classify slip events. This research develops a low-latency and computationally efficient vibration-based system to detect wheel-terrain slip events for skid-steer rovers with modest hardware requirements.
To this end, vibration datasets corresponding to various wheel-terrain slip values are generated. A Husky rover with two Inertial Measurement Unit sensors is used in indoor and outdoor test environments. Slip is induced at specific values by mechanically constraining the rover to reduce the Actual Rover Speed below the Commanded Rover Speed. The vibration datasets are used to train and validate a Support Vector Machine classifier to differentiate abnormally high slip events from normal low slip. The training is done with various sensor outputs, sampling time, and sampling frequency. The performance of the system is then evaluated in order to find which combinations of parameters are effective and to qualify the trade-offs in performance which come with less ideal parameter values
Integrating Advanced Sizing and Controllability Assessment Methods into an MDO Framework for Optimal Redundancy in UAV Design
Get PDFElectrically-powered multirotor Unmanned Aerial Vehicles ( UAV) are highly susceptible to rotor loss failures, which can result in catastrophic events in urban centers. Redundancy implementation can improve reliability, but heavily affects multirotor UAV performance. Hence, this research work aims at developing a design framework that may optimize multirotor UAVs for both performance and reliability. For this matter, a controllability assessment will be implemented in the design
of multirotor UAVs to ensure fault-tolerant design. An exploration of the design methodologies of multirotor UAVs demonstrates that only one
includes redundancy analysis linked to controllability. The methodologies are classified into three major groups: empirical, analytical modeling and simulations, and analytical catalog-based. These classes are compared in a case study to demonstrate that analytical modeling and simulations are best suited for redundancy implementation due to their affinity with both innovation and reliability. Since traditional controllability is not sufficient for multirotor UAVs, alternatives are evaluated. The Degree of Controllability ( DoC) is chosen since it possesses recovery time requirement potential and the simple inclusion of disturbances. After modification for the application of the DoC to multirotor UAVs, it is implemented within the optimization framework of an analytical modeling and simulation design methodology, optimizing multirotor UAVs for both performance and reliability. The potential recovery time requirements of the DoC prove inconclusive because its value scales with time. Future works, through reference recovery regions, could forge toward aircraft-level controllability requirements. Nevertheless, including the DoC as a constraint yields fault-tolerant designs at a severe cost in computing time
Reconciliation and Renewed Relationships in the Co-management of National Parks
Get PDFA new era of Indigenous-led collaborations signals a shift in approach by Parks Canada – in response to commitments to reconciliation – to the involvement of Indigenous peoples in the governance and management of national parks, national park reserves, and national marine conservation areas. However, co-management, the institutional arrangement on which these and other longstanding partnerships in parks contexts have been built, has a contested and uneven track record in meeting the needs, interests, and aspirations of Indigenous people. Using qualitative methods of governance analysis combined with interviews reflecting Indigenous and non-Indigenous perspectives, this thesis addresses the question: “what is the potential of co-management as a vehicle for reconciliation within national parks”? The thesis is comprised of two manuscripts. The first confronts a critical gap in empirical data about the content and context of formal national park co-management agreements through a scan of available agreements and the creation of a governance typology, as a basis for exploring strengths and weaknesses of agreement-making in serving reconciliation commitments. The second, through a community-partnered project with Vuntut Gwitchin First Nation, examines relationship-building processes in the context of Vuntut National Park, as an example of a mature claims-based northern national park co-management arrangement. Using the lens of ethical space, the research sheds light on enabling and constraining factors for relationship-building and offers insights into the principles and elements of an ethical space process for national park co-management arrangements supportive of Indigenous-state reconciliation. Overall, this thesis aims to contribute to understandings of the potential of co-management agreements to support reconciliation and renewed relationships in national parks
Early Wildfire Detection, Geolocation, and 3D Reconstruction Using Aerial Visible and Infrared Images
Get PDFIn this work, a novel unmanned aerial vehicle (UAV) system to detect and monitor early wildfire is proposed and verified with experimental flight tests on a DJI M300 quadrotor UAV. Several
strategies and algorithms are employed to fuse the on-board sensor information from the visible camera, infrared camera, inertial measurement unit (IMU), and global navigation satellite system
(GNSS). Developed strategies achieve the following functions: wildfire smoke and flame detection, wildfire spot geolocation, visible-infrared image registration, and wildfire local environment 3D reconstruction. For wildfire flame and smoke detection and segmentation, ResNet and attention-gate U-net are trained and deployed, which provide the semantic information for other modules in the proposed system. Simultaneous localization and mapping (SLAM) and multi-view stereo (MVS) algorithms are used to recover the camera poses and estimate the geolocation of the wildfire spot.
With the estimated geolocation information, a mode-based visible-infrared image registration algorithm is developed to decrease the false alarm by fusing the visible and infrared images. After that,
structure from motion (SfM) and MVS are utilized to recover the 3D scene of the local environment around the wildfire. Three independent outdoor experiments are conducted to verify the proposed detection, geolocation, registration, and local environment reconstruction algorithms. DJI M300 UAV together with a H20T camera are used as the platform in those flight testing experiments. The proposed system is proven to have promising applications in wildfire management
Can Beauty be a Catalyst for an Encounter with God?
Get PDFThis thesis leverages the theology of Hans Urs von Balthasar and others to explore beauty’s essential role in one’s spiritual life and how it can come to be viewed as central to one’s relationship with God. According to Balthasar, beauty, glory, and love are inseparably linked, and when beauty is restored to the rank of a transcendental, on par with truth and goodness, God's love can be experienced in a new and empowering way. The thesis explores the possibility that beauty is accessible to all and is not reserved for the few. The perception and experience of beauty is a skill that can be learned by training our physical and spiritual senses. The thesis explores how beauty can be found everywhere, in the microcosm and the macrocosm, in sadness and joy, in art and in nature. It also explores the possibility that, from a Christian perspective, the pinnacle of beauty, where it is most visible and most transformative, is in the incarnation of God in Jesus Christ and culminates in Christ’s kenotic self-giving on the cross
Application of GPU Accelerated Paired Explicit Runge-Kutta Methods to Turbulent Flow Over a Sphere
Get PDFThe design of next generation aircraft will rely on the use of high-fidelity computational fluid dynamics (CFD). For this purpose, the Paired Explicit Runge-Kutta (P-ERK) time stepping method was developed. It is a variation of explicit Runge-Kutta methods that allows the pairing of multiple methods within a given simulation. This results in higher stability methods being used in stiff regions, and lower cost methods being used in non-stiff regions, resulting in significant performance improvements. This work explores the utility of Graphical Processing Unit (GPU) acceleration combined with P-ERK schemes Subcritical flow over a smooth sphere at a Reynolds number (Re) of 3700 was used as the validation case. The flux reconstruction (FR) spatial discretization scheme was used with the implicit Large Eddy Simulation (ILES) turbulence modelling approach. Instantaneous quantities such as velocity fluctuations, Strouhal numbers, and time-averaged quantities such as drag coefficient, pressure coefficient, back pressure ratio, and Reynolds stresses were obtained. In addition, Q-criterion contours were generated and used to obtain separation angle and recirculation bubble length. This study shows that the P-ERK method can achieve good agreement with both reference simulation, and experimental data. Moreover, when compared to the traditional fourth order Runge-Kutta (RK) method the P-ERK scheme shows an average speedup factor of 4.73 using GPUs and of 5.82 using CPUs with regards to solution polynomial scaling, and with regards to resource scaling it requires approximately 8 times more resources using CPUs, or each CPU has an approximately 45 times greater runtime compared to one GPU. This is a significant reduction in computational times, while maintaining accuracy