Information

Abstract

Excellent explanations of feature visualization already exist in the form of interactive articles, e.g. DeepDream, Feature Visualization, The Building Blocks of Interpretability, Activation Atlas, Visualizing GoogLeNet Classes. They mostly rely on curated prerendered visualizations, additionally providing colab notebooks or public repositories allowing the reader to reproduce those results. While precalculated visualizations have many advantages (directability, more processing budget), they are always discretized samples of a continuous parameter space. In the spirit of Tensorflow Playground, this project aims at providing a fully interactive interface to some basic functionality of the originally Python-based Lucid library, roughly corresponding to the concepts presented in the “Feature Visualization" article. The user is invited to explore the effect of parameter changes in a playful way and without requiring any knowledge of programming, enabled by an implementation on top of TensorFlow.js. Live updates of the generated input image as well as feature map activations should give the user a visual intuition to the otherwise abstract optimization process. Further, this interface opens the domain of feature visualization to non-experts, as no scripting is required.

Additional Files and Images

Additional images and videos

video: Submission video for VISxAI workshop. video: Submission video for VISxAI workshop.

Additional files

Weblinks

  • VISxAI URL
    Explainable URL, accepted for the IEEE VIS 2019 Workshop on Visualization for AI Explainability.
  • VISxAI workshop
    2nd workshop on Visualization for AI Explainability in Vancover, Canada, October 21, 2019.
  • LucidPlayground GitHub
    Source code

BibTeX

@studentproject{sietzen-2019-wnn,
  title =      "Web-Interface for neural network feature visualization",
  author =     "Stefan Sietzen",
  year =       "2019",
  abstract =   "Excellent explanations of feature visualization already
               exist in the form of interactive articles, e.g. DeepDream,
               Feature Visualization, The Building Blocks of
               Interpretability, Activation Atlas, Visualizing GoogLeNet
               Classes. They mostly rely on curated prerendered
               visualizations, additionally providing colab notebooks or
               public repositories allowing the reader to reproduce those
               results. While precalculated visualizations have many
               advantages (directability, more processing budget), they are
               always discretized samples of a continuous parameter space.
               In the spirit of Tensorflow Playground, this project aims at
               providing a fully interactive interface to some basic
               functionality of the originally Python-based Lucid library,
               roughly corresponding to the concepts presented in the
               “Feature Visualization" article. The user is invited to
               explore the effect of parameter changes in a playful way and
               without requiring any knowledge of programming, enabled by
               an implementation on top of TensorFlow.js. Live updates of
               the generated input image as well as feature map activations
               should give the user a visual intuition to the otherwise
               abstract optimization process. Further, this interface opens
               the domain of feature visualization to non-experts, as no
               scripting is required.",
  month =      aug,
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2019/sietzen-2019-wnn/",
}