Bulletin of the American Physical Society
APS March Meeting 2022
Volume 67, Number 3
Monday–Friday, March 14–18, 2022; Chicago
Session Q32: Deep Learning Computer VisionRecordings Available
|
Hide Abstracts |
Sponsoring Units: GDS Chair: William Ratcliff, GDS Room: McCormick Place W-192B |
Wednesday, March 16, 2022 3:00PM - 3:12PM |
Q32.00001: A neural network can hear the shape of a drum Yueqi Zhao, Michael M Fogler We have developed an artificial neural network that reconstructs the shape of a polygonal domain given the first few dozen of its Laplacian (or Schrodinger) eigenvalues. This provides a practical answer to the famous “can one hear the shape of a drum” question posed by M. Kac in 1966. Having an encoder-decoder structure, the network maps input spectra to a latent space and then predicts the discretized image of the domain. We tested this network on randomly generated triangles, quadrangles, and pentagons. The prediction accuracy remains high for all of these shapes. All predictions scale according to the inputs under the Laplacian scaling rule. The network also recovers the continuous rotational degree of freedom beyond the symmetry of the lattice. The variation of the latent variables under the scaling transformation shows they are strongly correlated with the parameters of the domain (area, perimeter, and a certain function of the angles) from the Weyl expansion. |
Wednesday, March 16, 2022 3:12PM - 3:24PM |
Q32.00002: Deepfake Video Detection Using Biologically Inspired Geometric Deep Learning Steven Luo, Jansen Wong, Yash Agarwal, Aayush Sheth, Krish Jain, Eric Zhu, Eric Guan, Nilesh Chaturvedi, Pawel Polak Deepfakes are synthetic media in which a person’s appearance is altered to look like another’s. In order to identify them we trace the face by employing a 3D manifold obtained from Google’s open-source MediaPipe library. It provides us with a local topology of the face that is invariant to any homeomorphic transformation across consecutive frames. We measure the local muscle movement in the face using a heatmap of correlations between pixels in the corresponding frames. These unique biological characteristics are projected to the canonical face mesh and used as features in a 3D Convolutional Neural Network to quantify the authenticity of the face on the video. The proposed model is trained on publicly available databases of manipulated and real videos, including FaceForensics, FaceForensics++, and Deep Fakes Dataset, with more than 500 gigabytes of data available. |
Wednesday, March 16, 2022 3:24PM - 3:36PM |
Q32.00003: Unsupervised Machine Learning for Spatio-Temporal Characterization of Nanoscale Phenomena Imaged via Ultrafast Electron Microscopy Faran Zhou, Thomas E Gage, Haihua Liu, Ilke Arslan, Haidan Wen, Maria K Chan Advancements in microscopy techniques have made it possible to investigate dynamic structural phenomena at nanoscales. This work details the use of a machine learning based approach to extract quantitative information regarding the motion of features as captured by an ultrafast electron microscope (UEM). UEM is an emerging technique that uses pulsed electron beams to image structural dynamics at nanometer-picosecond resolutions. This spatio-temporal characteristic of a UEM dataset is one of the main challenges encountered during its analysis. Classical computer vision techniques for characterizing motion between image frames are parametric, and hence require manual supervision. In this work, a U-net type convolutional neural network is designed to take a pair of UEM images as input and generate the optical flow at each pixel as output. A custom loss function is defined, consisting of a photometric loss term and a gradient loss term. Additionally, the uncertainty associated with the estimate at each pixel is quantified using the Monte-Carlo Dropout method. The pixel-level motion computation provides a framework to correlate the phonon wavefront motion with the nanoscale interface structure characteristics, and this is demonstrated using FePS3 as an example. |
Wednesday, March 16, 2022 3:36PM - 3:48PM |
Q32.00004: Semantic Segmentation for Analysis of Melting of Nanoscale Ice via Fully Convolutional Neural Networks Arun Baskaran, Yulin Lin, Jianguo Wen, Maria K Chan Phase segmentation from electron microscopy datasets has emerged as a major subclass of computer vision problems for materials characterization. In this work, we show the application of semantic segmentation to image analysis of the melting of ice in the nanoscale regime, via in-situ observation by Transmission Electron Microscopy (TEM). Nano water confined into a carbon film based liquid cell were transferred into the ultra-high vacuum TEM chamber by a cryo-TEM holder, which provided the required control over temperature and phase. A high-speed K2 direct detection camera was used to record the melting process under low electron dose (<0.1 e-/nm2), at a rate of 400 frames per second. An ensemble of pre-trained U-Nets were trained to segment ice from the image frames. These neural networks were fine-tuned on a dataset consisting of image frames from the melting experiments. During inference, image frames from a given experiment are passed as input to the trained networks to generate time profiles of the area fraction occupied by ice. The demonstrated approach allows rapid feedback, with high segmentation accuracy and a measure to quantify the uncertainty which can be used to isolate low-quality images, and autonomous control of TEM experiments with high speed cameras. |
Wednesday, March 16, 2022 3:48PM - 4:00PM |
Q32.00005: One Visualization is Worth 1000 Words: Toward Automated Data Recovery and Interpretation from Past 3D Visualizations Laura E Brandt, William T Freeman Since the late 1700s, plots and graphs have been frequently used to visualize scientific data and convey results far more clearly than can be done with words. Unfortunately, much of that past work, even if surviving and converted to electronic form, is effectively inaccessible to most semantic queries. Unlike photographs and other pictorial presentations, plots and graphs are not interpretable by search engines. Further, even if a researcher identifies a figure relevant to their work, it is often non-trivial to recover the numerical data being represented. To address these challenges, researchers are actively working on methods for automatically extracting information from published 2D plots and figures to enable large-scale indexing. In this work, we turn our eye to 3D data visualizations. Using our recently-published SurfaceGrid dataset, we successfully train an artificial neural net to recover numerical data from 3D plots and graphical models using contour-based curvature cues that are widely used in published data visualizations. When calibration information is available, reconstructions have less than 0.5% mean-squared relative error. |
Wednesday, March 16, 2022 4:00PM - 4:12PM |
Q32.00006: Lithium Metal Battery Characterization using X-ray Imaging and Machine Learning Daniela Ushizima, Ying Huang, Jerome Quenum, David Perlmutter, Dilworth Parkinson, Iryna Zenyuk In an ever-demanding world for zero emission clean energy sources, vehicle electrification will bring major contributions as each clean car that substitutes one based on fossil fuel could save 1.5 tons of carbon dioxide per year. To expand the e-vehicle fleet, new solutions to store energy must deliver lighter, longer ranges, and more powerful energy batteries, such as solid-state lithium metal batteries (LMB). Different from traditional lithium-ion, LMB uses solid electrodes and electrolytes, providing superior electrochemical performance and high energy density. Some of the challenges of this new technology are to predict the cycling stability and to prevent the formation of lithium dendrite growth. This harmful phenomenon may occur during LMB charge and discharge, when lithium can deposit irregularly, building up dendrites (lithium plating) that leads to failures, such as short-circuit. These morphologies are key to the LMB quality, and they can be captured and analyzed using X-ray microtomography (XRT) scans. This presentation will show a new set of machine learning algorithms, and multiscale representation of XRT from LMB samples, that enable the quantification of LMB defects, as well as new protocols to monitor the lifespan of a LMB and the evolution of them during cycling. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700