April 25 Notes

In this lecture, we will explore how visualizations can be used beyond data exploration tasks, to understand something as complex as Machine Learning decision making.

A machine learning model derives the relationship between input and output using the computational units of its architecture.

Consider the below example, where an image is input into Resnet50, a popular image classification model and our task is to understand what led to the given decision.

The above image was classified as Broom (0.59% probability), Honeycomb (0.54% probability), Plunger (0.52% probability), Tennis Ball (0.51% probability) or Panpipe (0.46% probability).

HTML for today:

Visualizationthe decisions: Heatmap is one popular technique to explore important features.

Importance weights generated by the explanation algorithm

Dividing image into meaningful segments

Segments Created: 0