Day #185 - Interpretable Machine Learning for Computer Vision - Part II

Key Lessons

• Deep NN for Medical Decisions
• Deep NN for Understanding Scenes
• Deep NN for Visual Recognition

Interepretability of DNN

• Explanation for Algorithmic decision to deploy models
• Accuracy metrics, Feature space

Understanding CNN individual Units

• CNN Layers
• Sources of Data & Training

Deconvolution

• Visualize internal representations
• Back Propagation
• Back project Image Space

Gradient Based Visualization

• Iteratively use gradient to visualize / activate units
• Lower Layer capture textures
• Higher layer detect parts of objects

Data Driver Visualization

• Sample-based
• Visualize units at each layer
• Top Activated Images

Network Dissection

• Framework to quantify interpretability
• Intersection over union = Areas of Overlap / Area of Union
• RNN based Explanation Generator Model, Generate for classification models

Summary - Later Stage of CNN class selectivity higher

Talk #2 - Understanding Models via Visualization, Attribution and Semantic Identification

Key Summary

• Exploring the Deep Networks Black box
• Deep Network interepreted as a sequence of function

Generating iconic example

• Inverting Layers
• Reconstruct the image pixel by pixel

Attribution

• Find what parts of image are salient for deep network
• Finding Artifacts
• Sensitivity Analysis

Gradcam

• Propagating few layers 

Summary

Next List

Artificial Intelligence Imitation Learning - Tutorial - 2018 ICML
Faster R-CNN for Real-time Object Detection
Loss Functions for Regression and Classification
Tutorial on Generative adversarial networks - GANs as Learned Loss Functions
CS 188 | Introduction to Artificial Intelligence-Fall 2018

Happy Mastering DL!!!