skip to main content
US FlagAn official website of the United States government
dot gov icon
Official websites use .gov
A .gov website belongs to an official government organization in the United States.
https lock icon
Secure .gov websites use HTTPS
A lock ( lock ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

Explore Research Products in the PAR
It may take a few hours for recently added research products to appear in PAR search results.


Title: SplineCam: Exact Visualization and Characterization of Deep Network Geometry and Decision Boundaries
Current Deep Network (DN) visualization and inter-pretability methods rely heavily on data space visualizations such as scoring which dimensions of the data are responsible for their associated prediction or generating new data features or samples that best match a given DN unit or representation. In this paper, we go one step further by developing the first provably exact method for computing the geometry of a DN's mapping - including its decision boundary - over a specified region of the data space. By lever-aging the theory of Continuous Piece- Wise Linear (CPWL) spline DNs, SplineCam exactly computes a DN's geometry without resorting to approximations such as sampling or architecture simplification. SplineCam applies to any DN architecture based on CPWL activation nonlinearities, including (leaky) ReLU, absolute value, maxout, and max-pooling and can also be applied to regression DNs such as implicit neural representations. Beyond decision boundary visualization and characterization, SplineCam enables one to compare architectures, measure generalizability, and sample from the decision boundary on or off the data manifold. Project website: bit.ly/splinecam.  more » « less
Award ID(s):
1911094 1838177 1730574
PAR ID:
10466336
Author(s) / Creator(s):
; ; ;
Date Published:
Journal Name:
2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
Page Range / eLocation ID:
3789 to 3798
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. ; ; (, ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP))
    Computing or approximating the convex hull of a dataset plays a role in a wide range of applications, including economics, statistics, and physics, to name just a few. However, convex hull computation and approximation is exponentially complex, in terms of both memory and computation, as the ambient space dimension increases. In this paper, we propose DeepHull, a new convex hull approximation algorithm based on convex deep networks (DNs) with continuous piecewise-affine nonlinearities and nonnegative weights. The idea is that binary classification between true data samples and adversarially generated samples with such a DN naturally induces a polytope decision boundary that approximates the true data convex hull. A range of exploratory experiments demonstrates that DeepHull efficiently produces a meaningful convex hull approximation, even in a high-dimensional ambient space. 
    more » « less
  2. ; ; (, 75th APS-DFD November 2022)
    In this video, we show high-fidelity numerical results of supersonic spatially-developing turbulent boundary layers (SDTBL) under strong concave and concave curvatures and Mach = 2.86. The selected numerical tool is Direct Numerical Simulation (DNS) with high spatial/temporal resolution. The prescribed concave geometry is based on the experimental study by Donovan et al. (J. Fluid Mech., 259, 1-24, 1994). Turbulent inflow conditions are based on extracted data from a previous DNS over a flat plate (i.e., turbulence precursors). The comprehensive DNS information sheds important light on the transport phenomena inside turbulent boundary layers subject to strong deceleration or Adverse Pressure Gradient (APG) caused by concave walls as well as to strong acceleration or Favorable Pressure Gradient (FPG) caused by convex walls at different wall thermal conditions (i.e., cold, adiabatic and hot walls). In this opportunity, the selected scientific visualization tool is Virtual Reality (VR) by extracting vortex core iso-surfaces via the Q-criterion to convert them to a file format readable by the HTC Vive VR toolkit. The reader is invited to visit our Virtual Wind Tunnel (VWT) under a fully immersive environment for further details. The video is available at: https://gfm.aps.org/meetings/dfd-2022/6313a60c199e4c2da9a946bc 
    more » « less
  3. ; ; ; (, Preprints.org)
    The study presents an innovative pipeline for processing, compressing, and remotely visualizing large-scale numerical simulations of fluid dynamics in a virtual wind tunnel (VWT), leveraging Virtual and Augmented Reality (VR/AR) for enhanced analysis and high-end visualization. The workflow addresses the challenges of handling massive databases obtained via Direct Numerical Simulation (DNS) while maintaining visual fidelity, promoting full immersion, and ensuring efficient rendering for user interaction. We are performing fully immersive visualization of high-fidelity numerical results of supersonic spatially-developing turbulent boundary layers (SDTBL) under strong concave/convex curvatures at a freestream Mach number of 2.86 (i.e., supersonic flow). The selected numerical tool is Direct Numerical Simulation (DNS) with high spatial/temporal resolution. The comprehensive DNS information sheds important light on the transport phenomena inside turbulent boundary layers subject to strong deceleration or Adverse Pressure Gradient (APG) caused by concave walls as well as to strong acceleration or Favorable Pressure Gradient (FPG) caused by convex walls at different wall thermal conditions (i.e., Cold, Adiabatic and Hot walls). The process begins with .vts file input from DNS, which is visualized using the ParaView software. Multiple iso-contours for parameters such as velocity and temperature are generated, applying custom formulas to create visualizations at various floating-point precisions (16-bit, 32-bit, 64-bit). These visualizations, representing different fluid behaviors based on DNS with high spatial/temporal resolution and millions of “numerical sensors”, are treated as individual time frames and exported in GLTF format. Our approach demonstrates significant improvements in rendering speed and user experience, particularly when dealing with datasets comprising hundreds of high-resolution frames from Computational Fluid Dynamics (CFD) simulations. By utilizing server-side compression and cloud rendering, we overcome the limitations of on-device processing, enabling smooth and responsive interactions even with large, complex fluid dynamics datasets. This pipeline represents a substantial advancement in scientific visualization of fluid dynamics, offering researchers and engineers a powerful tool for exploring and analyzing large-scale CFD simulations in an immersive, intuitive environment. Additionally, we leverage Unity’s Profile Analyzer and Memory Profiling tools with the purpose of identifying major bottlenecks and resource-consuming events during contour running, with a keen focus on enhancing GPU and CPU efficiency. In conclusion, the materials and methods employed in this project were instrumental in systematically collecting, analyzing, and interpreting performance data from DNS databases. Future work will focus on optimizing compression algorithms for fluid-specific data and expanding the range of supported simulation parameters to enhance the pipeline’s versatility across various fluid dynamics applications. 
    more » « less
  4. ; ; ; ; ; (, Bioinspiration & Biomimetics)
    Pinniped vibrissae possess a unique and complex three-dimensional topography, which has beneficial fluid flow characteristics such as substantial reductions in drag, lift, and vortex induced vibration. To understand and leverage these effects, the downstream vortex dynamics must be studied. Dye visualization is a traditional qualitative method of capturing these downstream effects, specifically in comparative biological investigations where complex equipment can be prohibitive. High-fidelity numerical simulations or experimental particle image velocimetry are commonplace for quantitative high-resolution flow measurements, but are computationally expensive, require costly equipment, and can have limited measurement windows. This study establishes a method for extracting quantitative data from standard dye visualization experiments on seal whisker geometries by leveraging novel but intuitive computer vision techniques, which maintain simplicity and an advantageous large experimental viewing window while automating the extraction of vortex frequency, position, and advection. Results are compared to direct numerical simulation (DNS) data for comparable geometries. Power spectra and Strouhal numbers show consistent behavior between methods for a Reynolds number of 500, with minima at the canonical geometry wavelength of 3.43 and a peak frequency of 0.2 for a Reynolds number of 250. The vortex tracking reveals a clear increase in velocity from roll-up to 3.5 whisker diameters downstream, with a strong overlap with the DNS data but shows steady results beyond the limited DNS window. This investigation provides insight into a valuable bio-inspired engineering model while advancing an analytical methodology that can readily be applied to a broad range of comparative biological studies. 
    more » « less
  5. ; ; ; ; ; (, bioRxiv)
    ABSTRACT Synapses are often precisely organized on dendritic arbors, yet the role of synaptic topography in dendritic integration remains poorly understood. Utilizing electron microscopy (EM) connectomics we investigate synaptic topography inDrosophila melanogasterlooming circuits, focusing on retinotopically tuned visual projection neurons (VPNs) that synapse onto descending neurons (DNs). Synapses of a given VPN type project to non-overlapping regions on DN dendrites. Within these spatially constrained clusters, synapses are not retinotopically organized, but instead adopt near random distributions. To investigate how this organization strategy impacts DN integration, we developed multicompartment models of DNs fitted to experimental data and using precise EM morphologies and synapse locations. We find that DN dendrite morphologies normalize EPSP amplitudes of individual synaptic inputs and that near random distributions of synapses ensure linear encoding of synapse numbers from individual VPNs. These findings illuminate how synaptic topography influences dendritic integration and suggest that linear encoding of synapse numbers may be a default strategy established through connectivity and passive neuron properties, upon which active properties and plasticity can then tune as needed. 
    more » « less
  • Citation Formats
  • MLA
  • APA
  • Chicago
  • BibTeX
  • Save / Share this Record
  • Send to Email