Revision as of 22:02, 6 October 2020

DSC 180 Capstone Domain: Explainable AI (Section A01)

• Instructor: Jurgen Schulze
• Discussion: Wednesdays 12-12:50pm on Zoom at https://ucsd.zoom.us/j/9100475160
• Piazza Discussion Board: https://piazza.com/ucsd/fall2020/dsc180

Overview

In this capstone domain we are going to study how we can make machine learning systems more user friendly by exploiting additional knowledge we can derive from the system and present it to the user. These types of systems are called Explainable AI.

The example we are going to use in this class is object recognition in images. We are first going to go over the basics of CNNs, then learn about saliency and attention maps, and finally we are going to implement the Grad-CAM algorithm in PyTorch and apply it to the COCO image data set.

The quarter will end with a proposal for your capstone project, which you will be working on in the winter quarter.

This class will be entirely remote.

Schedule

Papers

• Learning Deep Features for Discriminative Localization
• Grad-CAM: Why did you say that?Visual Explanations from Deep Networks via Gradient-based Localization
• Visualizing CNNs with deconvolution

Useful Links

• Main DSC Capstone Website
• Introduction to how CNNs Work
• Tensorflow Playground
• Grad-CAM implementation in Pytorch
• COCO Image Dataset

• What CNN layer outputs look like in Tensorflow:
  • https://github.com/conan7882/CNN-Visualization
  • https://medium.com/@awjuliani/visualizing-neural-network-layer-activation-tensorflow-tutorial-d45f8bf7bbc4

• Interesting visualization: http://people.csail.mit.edu/torralba/research/drawCNN/drawNet.html
• More stuff:
  • https://towardsdatascience.com/how-to-visualize-convolutional-features-in-40-lines-of-code-70b7d87b0030
  • https://towardsdatascience.com/understanding-your-convolution-network-with-visualizations-a4883441533b
  • http://cs231n.github.io/understanding-cnn/
  • http://cs-people.bu.edu/jmzhang/excitationbp.html