Dark, Beyond Deep: A Paradigm Shift to Cognitive AI with Humanlike Common Sense is a paper by Zhu, et al. (2020) I discussed in Dr. Antonio Vergari’s weekly lab meeting.

My slides are here.

You can try the Heider-Simmel experiment yourself.

Here is a live demo of the Spatio-Temporal Causal And-Or Graph (STC-AoG) in action.

The thumbnail is cropped from the book by Siyuan, Huang and Zhu, Song-Chun on Computer Vision: Stochastic Grammars for Parsing Objects, Scenes, and Events

These were my presentation notes:

  • AI is not just about problem solving, but also understanding.
  • Understanding leads to:
    • Better generalization
    • Small data for big tasks
    • Although small data means more expert architectural engineering of the machine learning model
    • Understanding a problem is more humanlike
  • Task-driven vision
  • Top-down and bottom-up parsing of a sequence of images.
  • How would an And-Or graph not overfit to a room’s layout?