Browsing by Subject "Neural networks"

Now showing 1 - 9 of 9
  • Thumbnail Image
    Item
    Advanced analysis and background techniques for the cryogenic dark matter search.
    (2010-01) Qiu, Xinjie
    The Cryogenic Dark Matter Search (CDMS) used Ge and Si detectors, operated at 50mK, to look for Weakly Interacting Massive Particles (WIMPs) which may make up most of the dark matter in our universe. This dissertation describes the simulation, analysis, and results of the first WIMP-search data runs of the CDMS experiment between October 2006 and July 2007 with its 5 Towers of detectors at the Soudan Underground Laboratory. A blind analysis, incorporating improved techniques for event reconstruction and data quality monitoring, resulted in zero observed events. The results of this work place the most stringent limits yet set upon the WIMP-nucleon spin-independent cross section for WIMP masses above 44 GeV/c2, as well as setting competitive limits on spin-dependent WIMP-nucleon interactions.
  • Thumbnail Image
    Item
    Development of On-Line Control Strategies in Freeway Networks, Phase 2: Final Report
    (Minnesota Department of Transportation, 1998-05) Stephanedes, Yorgos J.; Liu, Xiao; Liu, Lu; Michel, Bernard R.
    Most traffic-responsive freeway ramp metering systems select metering rates from predetermined rate libraries. The efficiency of such systems is impaired by the lack of an efficient analysis tool that can evaluate and update the thresholds and rate libraries used by the meter controllers. In this project, a control-emulation method is developed to evaluate various automatic rateselection strategies; the new modeling features of this system are described in detail. Various rate selection strategies (based on neural network processing, exit ramp volume, and real time bottleneck/dynamic zone determination) are described and evaluated in comparison with the current Minneapolis-St. Paul strategy. An online traffic volume predictor based on Kalman filtering is developed, and integrated into the control-emulation module. A simulated annealing optimization algorithm, previously implemented on a supercomputer, is re-implemented on a personal computer and integrated into the simulation module.
  • Thumbnail Image
    Item
    Finger Movement Classification via Machine Learning using EMG Armband for 3D Printed Robotic Hand
    (2019-09) Bhatti, Shayan Ali
    Millions of people lose their limbs due to accidents, infections and/or wars. While prosthetics are the best solution for amputees, designing autonomous prosthetic hand that can perform major operations is a complicated task and thus the prosthetic hands that are designed are very expensive and also a bit heavy. The biggest challenge in designing a prosthetic hand is the classification of EMG signals generated by neurons in the arm to distinguish finger movements. These EMG signals vary in strength from person to person and from movement to movement. This thesis proposes a computationally efficient way that uses Machine Learning to classify 5 and 12 finger movements from EMG signals captured by a device called “Myo Gesture Control Armband”. Further, an ergonomic design of robotic hand is also presented that is small, lightweight and cheap, designed using a 3D printer.
  • Thumbnail Image
    Item
    Image Enhancement Algorithms to Improve Robustness and Accuracy of Pre-trained Neural Networks for Autonomous Driving
    (2023-01) Joshi, Himanshu
    This research proposes a generalized data-driven approach for improving the pre-training image datasets fed to neural networks(NNs). The algorithms developed and tested in this work could substantially enhance the image bringing out the critical spatial information in the image for better NN performance. This image enhancement technique consists of two main components: image colorization and contrast enhancement. Image colorization is implemented to obtain a color-corrected image from a grayscale image. The traditional global contrast enhancement algorithm is extended to Smoothened Variable Local Dynamic Contrast Improvement (SVL-DCI) to boost local contrasts within an image frame that suffers from under/over-exposed lighting conditions. SVL-DCI algorithm is developed and thoroughly tested in the present thesis that could run in real-time as a pre-training algorithm for NNs. We implemented SVL-DCI on the 3P lab dataset of 2470 images and observed competitive improvement in the performance of the investigated NNs for object recognition and lane detection.
  • Thumbnail Image
    Item
    Object tracking in aerial video of smoky environments
    (University of Minnesota. Institute for Mathematics and Its Applications, 2011-02) Rivera, Mariano; Fernandes, Praphat Xavier; Hernández, Francisco; Martínez, Hugo; McDonald, Matt; Ohlmacher, Scott W.; Wiens, Jeffery
  • Thumbnail Image
    Item
    Physics-Based Artificial Intelligence Models for Vehicle Emissions Prediction
    (2021-05) Panneer Selvam, Harish
    On-board diagnostics (OBD) data contain valuable information including real-world measurements of vehicle powertrain parameters. These data can be used to gain a richer data-driven understanding of complex physical phenomena like emissions formation during combustion. In this thesis, a physics-based artificial intelligence framework is developed to predict and analyze trends in engine-out NOx emissions of diesel and diesel-hybrid heavy-duty vehicles. This framework differs from black box machine learning models presented in previous literature because it incorporates engine combustion parameters that allow physical interpretation of the results. Based on chemical kinetics and the characteristics of diffusive combustion, NOx emissions from compression ignition engines primarily depend non-linearly on three parameters: adiabatic flame temperature, intake oxygen concentration, and combustion time duration. Here, these parameters were calculated from available OBD data. Non-linear regression coupled with a novel Divergent Window Co-occurrence (DWC) Pattern Detection algorithm is observed to be an effective method to predict NOx emissions and analyze driving patterns from the OBD data where prediction errors are high. The proposed framework is validated for generalizability with a separate vehicle OBD dataset, a sensitivity analysis is performed on the prediction model, and its predicted values are compared with that from a black-box deep neural network. The results show that NOx emissions predictions using the proposed model has around 55% better root mean square error, and around 60% higher mean absolute error compared to the baseline NOx prediction model from previously published work. This framework serves as a transparent and interpretable physics-based model to predict NOx emissions using OBD data as input. Furthermore, linearizing the physics-based NOx equation provides an opportunity to evaluate several machine learning regression techniques. The results show that an ensemble learning bagging-type model like random forest regression (RFR) is highly effective in predicting engine-out NOx emissions. We also show that real-world OBD data has high heterogeneity with clustered co-occurrences of vehicle parameters. In terms of accuracy, the developed RFR model provides an average of 53% improvement in R2 value and 42% better mean absolute error (MAE) for NOx emissions predictions compared to non-linear regression models, and provided the opportunity to interpret the results because of its linkage to physical parameters. We also perform a feature importance analysis for the RFR Model and compare prediction results with black box deep neural network and non-linear regression models. Based on its high accuracy and interpretability, the developed RFR model has potential for use in on-board NOx prediction in engines of varying displacement and design.
  • Thumbnail Image
    Item
    Real-Time Prediction of Freeway Occupancy for Congestion Control
    (Center for Transportation Studies, University of Minnesota, 1997-09) Cherkassky, Vladimir; Yi, Sangkug
    Accurate traffic prediction is critical for effective control of on-ramp traffic (ramp metering). During congestion, traffic shock waves propagate back and forth between the detectors, and traffic becomes inherently non-stationary and difficult to predict. Recently, several adaptive non-linear time series prediction methods have been developed in statistics and in artificial neural networks. We applied these methods to develop real-time prediction of freeway occupancy during congestion periods, from current and time-lagged observations of occupancy at several (neighboring) detector stations. This study used the following function estimation methodologies for real-time occupancy prediction: two statistical techniques, multivariate adaptive regression splines (MARS) and projection pursuit regression; two neural network methods, multi-layer perceptrons (MLP) and constrained topological mapping (CTM). All these methods were applied to freeway occupancy data collected on I-35W during morning rush hours. Data collected on one day was used for training (model estimation), whereas the data collected on a different day was used for testing, i.e., estimating the quality of prediction (generalization). Results for this study indicate that the proposed methodology provides 10-15% more accurate prediction of traffic during congestion periods than the approach currently used by Minnesota DOT.
  • No Thumbnail Available
    Item
    Supporting Data for "A novel machine learning method for accelerated modeling of the downwelling irradiance field in the upper ocean"
    (2022-04-27) Hao, Xuanting; Shen, Lian; haoxx081@umn.edu; Hao, Xuanting; University of Minnesota Fluid Mechanics Lab
    The training data are generated from the Monte Carlo simulation of oceanic irradiance field. They can be used for training a neural network that significantly accelerates the prediction of irradiance in the upper ocean.
  • Thumbnail Image
    Item
    Towards Hardware-Software Co-design for Energy-Efficient Deep Learning
    (2023-06) Unnikrishnan, Nanda
    Artificial intelligence (AI) has become an increasingly important and prevalent technology in today’s world. The past decade has seen tremendous growth in AI with it being used in a wide range of applications, including healthcare, finance, transportation, research, manufacturing, and even entertainment. One of the most significant advancements in AI has been the development of deep neural networks (DNNs), which have revolutionized the field by providing unprecedented human-like performance in solving many real-world problems. However, the computations involved in DNNs are expensive and time-consuming, especially for large and complex networks. Additionally, a variety of models, like convolutional neural networks (CNNs), recurrent neural networks (RNNs), transformers, and graph neural networks (GNNs), pose significant challenges for hardware design, particularly due to the diverse set of operations used. Each operation brings its own set of challenges for energy, performance, and memory that do not always align with one another precluding a one size fits all solution. The thesis addresses the above challenges in three parts. The first part tries to develop a fundamental understanding of the different operations involved in different DNN models. This thesis explores the evolution of brain-inspired computing models from a historical context, focusing on DNNs, CNNs, RNNs, and GNNs among others. This provides the necessary context for optimizing DNN operations for training and inference. The second part of the thesis proposed hardware-software co-design techniques inspired by the design of DSP systems to address energy, computation, and memory challenges during training for CNNs. The thesis proposes a novel approach for using systolic architectures to train convolutional neural networks using gradient interleaving, called InterGrad. The approach involves interleaving the computations of two gradients on the same configurable systolic array, resulting in significant savings in terms of the number of cycles and memory accesses. The proposed method uses 25% fewer cycles and memory accesses, and 16% less energy in state-of-the-art CNNs, and up to 2.2× fewer cycles and memory accesses in the fully connected layers. The thesis also presents a novel optimization approach called LayerPipe, which explores how to partition optimally and pipeline DNN training workload on multi-processor systems. LayerPipe can better balance workloads while minimizing the communication overhead. LayerPipe achieves an average speedup of 25% and upwards of 80% with 7 to 9 processors when compared to prior approaches such as PipeDream. Lastly, the thesis explores the design of dedicated hardware accelerators for graph neural networks (GNNs). The proposed SCV-GNN method uses a novel sparse compressed vectors (SCV) format optimized for the aggregation operation. The proposed method achieves a geometric mean speedup of 7.96× and 7.04× over a compressed sparse column (CSC) and compressed sparse rows (CSR) aggregation operations, respectively, and reduces the memory traffic by a factor of 3.29× and 4.37× over CSC and CSR, respectively.