Concepts and tools for fairness, explainability, and robustness in machine learning

Description

Ilya Feige details the AI safety concerns—fairness, robustness, and explainability—that are already relevant for ML models in use today.

Fairness ensures that a model’s predictions do not unethically discriminate against protected groups. This is a subtle matter, since unfair biases can exist in the data used to train the model as well as in the model’s own decision-making algorithm; moreover, precise definitions of such unfair biases are often mathematically contradictory and not widely agreed upon.

Robustness builds expectations for how an ML model will behave upon deployment in the real world. It addresses the questions of estimating uncertainties in its predictions and whether or not the model is robust to perturbed data.

Explainability tackles the question of how an ML model makes its predictions. An improved understanding of the model is essential for detecting, avoiding, and removing its failure modes; for earning public trust in the algorithm; and for introducing effective policies to regulate the technology.

Ilya demonstrates the AI safety tools developed at Faculty so you gain a detailed understanding of its models from these three perspectives. You’ll leave with the knowledge to deploy fair, robust, and explainable models, with a focus on concepts and examples.