Abstract : Considerable increases in GPU and high-performance computing have led to significant advances in machine learning methodologies. This mini-symposium will focus on recent advances in methodologies and applications of deep and machine learning to analyze two types of data. In the first session, our speakers will discuss recent advances in machine learning based methods for predicting time series data. The second session will focus on deep learning for analyzing and interpreting imaging data. Mathematical topics of this mini symposium include, but are not limited to, physics-driven machine learning, deep learning, reinforcement learning, computer vision, and optimal transport.
Organizer(s) : Kevin Flores, Ryan Murray, Hien Tran
[04949] Introduction to miniymposium session: Applications of machine learning to analyzing time-series and imaging data
Format : Talk at Waseda University
Author(s) :
Kevin Flores (NC State University)
Abstract : This talk is an introduction to the minisymposium session on "Applications of machine learning to analyzing time-series and imaging data".
[05230] Reinforcement Learning in a Digital Twin Framework for the Stabilization of an Inverted Pendulum
Format : Talk at Waseda University
Author(s) :
Hien Tran (North Carolina State University)
Abstract : In this talk we benchmark common reinforcement learning algorithms on a modified version of OpenAI Gym's Cartpole: a virtual environment simulating the dynamics of an inverted pendulum. The reinforcement learning algorithms that we used to stabilize the virtual inverted pendulum included the Policy Gradient, Actor-Critic, and Proximal Policy Optimization. We then transferred the trained neural network models from the virtual environment to the real physical inverted pendulum to verify their performances. While all of the reinforcement learning algorithms were able to satisfactorily balance the real inverted pendulum, Actor-Critic is best able to adequately reject disturbances.
[03788] Predicting Bladder Pressure and Contractions from Non-Invasive Time-Series Data
Format : Talk at Waseda University
Author(s) :
Erica M Rutter (University of California, Merced)
Abstract : Symptoms of bladder dysfunction can be alleviated by electrical stimulation of nerves at the start of a contraction. However, determining when a bladder contraction will occur remains an active area of research. Due to the extremely dense time-series data, we employ statistical and machine learning methods to predict bladder pressure from external nerve data. These bladder pressures are used to predict the onset of bladder contractions with high sensitivity and specificity.
[04294] Leveraging topological data analysis for parameter estimation of an agent-based model of collective motion
Format : Talk at Waseda University
Author(s) :
Kyle Nguyen (North Carolina State University)
Carter Jameson (North Carolina State University)
John Nardini (The College of New Jersey)
Kevin Flores (North Carolina State University)
Abstract : Understanding the social interaction between members of groups in the context of collective motion is significant to gain the insights on the link between local and global behaviors. By leveraging topological data analysis, we use dimensional reduction techniques on topological features to visually cluster different time series of collective motion simulations of an agent-based model into groups. We also propose inverse problem approaches for parameter estimation of this particular agent-based model of collective motion.
[04425] Few-Shot Learning for Leaf and Vein Segmentation
Format : Talk at Waseda University
Author(s) :
John Lagergren (Oak Ridge National Laboratory)
Abstract : Plant phenotyping is a primary bottleneck in understanding plant adaptation and the genetic architectures underlying complex traits at population scale. We address this challenge by leveraging few-shot learning with convolutional neural networks (CNNs) to segment the leaf body and visible venation of P. trichocarpa leaf images obtained in the field. Biological traits are extracted from the resulting segmentations, validated using real-world measurements, and used to conduct a genome-wide analysis to identify genes controlling the traits.
[04513] Analysis of spatial transcriptomics using deep learning and optimal transport
Format : Talk at Waseda University
Author(s) :
Zixuan Cang (North Carolina State University)
Abstract : The emerging single-cell and spatial genomics techniques allow us to elucidate the governing rules of multicellular systems with unprecedented resolution and depth. These datasets are often high-dimensional, complex, and heterogeneous. Mathematical tools are needed to extract biological insights from such data. In this talk, we will discuss several computational methods for exploring tissue structures, temporal signatures, and cell-cell communication processes on spatial transcriptomics data as well as supervised optimal transport motivated by the biological applications.
[04520] Applied Machine Learning for Overhead Imagery
Format : Talk at Waseda University
Author(s) :
Adam Attarian (Pacific Northwest National Laboratory)
Abstract : Applying machine learning techniques to collected overhead imagery presents many challenges not normally encountered in traditional object detection and classification problems. Complex sensing geometries, small target size, and lack of sufficient training data are all problems that must be mitigated. In this talk, we provide an overview of machine learning approaches and techniques to derive meaningful information from overhead imagery.
[05238] Solving inverse and forward problems in the water quality model by neutral networks
Format : Talk at Waseda University
Author(s) :
Quy Muoi Pham (The University of Danang - University of Science and Education)
Abstract : In this talk, we study the forward and inverse problems in BOD-DO models and present the Physic Inform Neural Network method to solve these problems. We first introduce the fully deep neural network and some well-known results about the approximation of fully deep neural networks to functions of classes. Then, we present the Physic Inform Neural Network method to solve the forward and inverse problems in BOD-DO models. We apply the method to solve some specific numerical examples. The method can be generalized for complex river quality models, e.g., 2D or 3D BOD-DO models or river quality models with more than two indicators. For complex river systems, we can use segmentation techniques to divide the river into some segments and in each segment, we can use the proposed method to solve the forward and inverse problems.
