AI/ML+Physics Part 3: Designing an Architecture - presented by Prof. Steve Brunton

AI/ML+Physics Part 3: Designing an Architecture

Prof. Steve Brunton

Prof. Steve Brunton
Slide at 24:26
U -NET - ARCHITECTURE BASIS OF DIFFUSION MODELS SEGMENTATION SUPER RESOLUTION DIFFUSION MODELS input image output tile segmentation 128 128 256 128 conv 3x3, ReLU copy and crop 1024 max pool 2x2 1024 up-conv 2x2 conv 1x1 Ronneberger, Fischer, Brox, arxiv, 2015
1
References
  • 1.
    O. Ronneberger et al. (2015) U-Net: Convolutional Networks for Biomedical Image Segmentation.
Share slide
Summary (AI generated)

The basis of many diffusion models includes an inductive implicit bias. The structure of these models highlights the multiscale nature of observations in the real world, both in space and time. When looking at a picture of the real world, this multiscale structure is evident. This architecture is adept at parameterizing natural images, scenes of traffic, cities, and other similar objects.