Day #184 - Interpretable Machine Learning for Computer Vision - Part I

Part I - Key Lessons

• ML is powerful hammer to solve lot of problems
• Complicated for humans to understand
• Building Linear classifier on hundred features

Burning Questions

• Decision Trees are interpretable
• As Data point grows it becomes even more difficult
• Rule list approach / Rule set approach
• No One size fits all method

• Reflect Fairness
• Reflect Experience

Where you don't need interepretability

• No Significant consequences
• Sufficiently studied problem

Interpretability Methods

• Before building model we can improve interpretability
• Other options - While bulding or after its built

Interepretability Options

• Before Building Model - Exploratory Data Analysis - Visualization - Exploratory Analysis - Mean, STD, K-Means, KNN - Facets - Visualization Tool
• Building a New Model - What is medium and constraints we use to explain - Rules, Examples, Sparsity and Monotonicity, Learn function for target variable
• After building a model - Ablation test, Input-feature importance, Concept importance 

Ablation test - Train without feature and see the impact
Fit Linear functions - Put Approximations / first derivatives
Local explanations - Fitting function locally true (About one data point)
Experiment - Saliency maps look similar after randomizing weights (Paper - Sanity Checks for Saliency Maps)

Evaluate Interpretability Methods

• Do human experiments
• Formulate experiment where you have ground truth

t-SNE - Understand Vision Models

• PCA Representation of Pixel values (Maximum variations captured)
• PCA learns linear mapping, preserve large pairwise distances
• KL divergence preserves the local data structure

Next Part II

Happy Mastering DL!!!