Black Box Algorithms

Chris Olah on What the Hell is Going on Inside Neural Networks

by

Author: Robert Wiblin, Keiran Harris, Chris Olah; Publisher: 80,000 Hours; Publication Year: 2021. The following podcast episode features Chris Olah, a researcher at Google focused on data interpretability, specifically as it relates to neural networks and big machine learning algorithms. The interview provides a deep dive into some of the issues around data interpretability and why the data “black box” issue is a problem worth solving…

fairmodels: a Flexible Tool for Bias Detection, Visualization, and Mitigation in Binary Classification Models

by

Author: Jakub Wisniewski, Przemyslaw Biecek; Publisher: The R Journal; Publication Year: 2022. The following article discusses how as more sophisticated machine learning methods become more ubiquitous, a culture of classifying between “explainable” and “unexplainable” models has become the norm. Researchers at MI2 DataLab have developed an opensource R package which provides a convenient and flexible workflow for detecting…

Manipulating SHAP via Adversarial Data Perturbations (Student Abstract)

by

Author: Hubert Baniecki, Przemyslaw Biecek; Publisher: Proceedings of the AAAI Conference on Artificial Intelligence; Publication Year: 2022. The following article discusses how Shapley Additive Explanations (SHAP) values are a popular metric for identifying variable importance for black-box machine learning models. This paper demonstrates that despite it’s popularity, the SHAP values can be manipulated using a genetic algorithm to increase or decrease variable importance by…

What is “Data Ethics”, and Why is it Important

by

Author: Dennis Hirsch; Publisher: Centre for Information Policy Leadership; Publication Year: 2021. The following article discuses what exactly is data ethics. Dennis Hirsch, a lawyer from Ohio State, asks this question in a study conducted in late 2021. In his study, he interviewed more than 20 data scientists and surveyed 50 more to gather information on data ethics from different perspectives. He found that, in its current iteration, it has mostly…

The Alignment Problem: Machine Learning and Human Values

by

Author: Brian Christian; Publisher: W. W. Norton & Company; Publication Year: 2020. The following book focuses on word2vec, an algorithm from 2013 that allows for vector computation of words. That is, word2vec allows for computations like king – man + woman = queen. Google used word2vec as part of Search and Translate. Multiple hiring platforms also used the algorithm. Researchers eventually…