owned this note   changed 6 years ago  
Linked with GitHub

Using Lucid to Visualize Neural Networks - Yufeng Guo

歡迎來到 PyCon TW 2019 共筆

Image Not Showing Possible Reasons
  • The image file may be corrupted
  • The server hosting the image is unavailable
  • The image path is incorrect
  • The image format is not supported
Learn More →

共筆入口：https://hackmd.io/@pycontw/2019
手機版請點選上方 按鈕展開議程列表。

PyCon Taiwan 2019 PyNight 活動規劃與 BoF 簽到/建立交流主題
https://hackmd.io/NaI9ymzRQ5urCvwOQ31WAA

從這開始

Optimization "Objectives": What activations are we optimizing for?

Diversity through dataset examples

  • "Square"-like grid shapes vs. circular stripings
  • Curvy

Shortcomings: Does not always make sense

  • Perhaps neurons aren't the right semantic units for understanding neural units
  • Random Combinations
  • Hand-crafted Combinations
  • Interpolating between 2 Neurons

Lucid

  • Infrastructure and tools for research in NN

Structure Dictionary

  • Activation: a combination of many neurons

Detection Across Layers

  • Edge in earlier layers
  • More sophisticated shapes and object parts in later layers

Scale Images by Activation Magnitude

  • How does information propagate through the layers?
  • Where are the important portions of the image?

Spatial Attritubes with Saliency Maps

Activation Athlases

  • Individual Neurons \(\rightarrow\) Parewise Interactions \(\rightarrow\) Spatial Activations \(\rightarrow\) Activation Atlas
  • Compute activation grid for each of your (1 million) images
  • UMAP (Uniform Manifold Approximation and Projection) to project down to 2D
  • Draw a grid over the projection, and averge the activations per grid
  • (Optional) Size the grid cells according to the density

Understand one specific classification

  • E.g., Fireboat vs. Streetcar

Resources

tags: PyConTW2019
Select a repo