Explore a comprehensive lecture on unmasking clever Hans predictors and assessing machine learning outcomes. Delve into popular heatmap techniques, Layer-wise Relevance Propagation (LRP), and its characteristics. Learn about LRP Taylor Series Expansion, backpropagation of last layer relevance, and LRP rules. Compare LRP heatmaps to Sensitivity Analysis for accuracy. Discover SPRAY (Spectral Relevance Analysis) and its application in cracking Fisher Vectors vs DNN predictions. Examine aggregate heatmap visualizations for various objects and uncover potential faults in DNN predictions. Gain valuable insights into assessing what machines truly learn and how to interpret their decision-making processes.
Unmasking Clever Hans Predictors and Assessing Machine Learning - Spring 2021
University of Central Florida via YouTube
Overview
Syllabus
Intro
Overview
Clever Hans (Horse)
Popular Heatmap Techniques
Layer wise Relevance Propagation
Characteristics of LRP
LRP Taylor Series Expansion for decomposition for last layer
Backpropagating Last Layer Relevance
LRP Rules
LRP Taylor Series Result LRP Heatmaps are more accurate sa compared to Sensitivity Analysis
Where & how to use LRP
SPRAY (Spectral Relevance Analysis) (Apply Spectral Cluster on LRP)
Spectral Clustering
Spectral Analysis Flow
Fisher Vector Model
SPRAY In Action, Cracking Fisher Vectors Vs DNN predictions
Aggregate Heat Map visualizations for boat & horse
SPRAY In Action, The Aeroplane Fault • Aeroplane is detected based on its background in both models
SPRAY In Action, DNN with Aeroplane Fault
Conclusion & Thoughts
Taught by
UCF CRCV
