Adversarial Examples, Protein Folding, and Shapley Values
Journal Club
English - April 28, 2020 20:28 - 45 minutes - 52.6 MBMathematics Science Education computerscience machinelearning modelinterpretability Homepage Download Google Podcasts Overcast Castro Pocket Casts RSS feed
George dives into his blog post experimenting with Scott Lundberg's SHAP library. By training an XGBoost model on a dataset about academic attainment and alcohol consumption can we develop a global interpretation of the underlying relationships? Lan leads the discussion of the paper Adversarial Examples Are Not Bugs, They Are Features by Ilyas and colleagues. This papers proposes a new perspective on adversarial susceptibility of machine learning models by teasing apart the 'robust' and the 'non-robust' features in a dataset. The authors summarizes the key take away message as "Adversarial vulnerability is a direct result of the models’ sensitivity to well-generalizing, ‘non-robust’ features in the data." Last but not least, Kyle discusses Alphafold!