Can an AI model anticipate how well it will perform in the wild? > In many important applications, AI models are trained on labeled data but when deployed in the wild, labels are not readily available (for example in medical imaging where the model is identifying a cancerous patch, "ground-truth" labels may require expert examination). A critical question is -- in
adversarial Featured Robust Attribution Regularization Recent work on training neural networks to have robust attributions, to improve their trustworthiness and resilience to adversarial attacks.
deep learning Probing Deep Network Behavior with Internal Influence In recent work with my colleagues at CMU, we focus on bringing greater transparency to these mysterious, yet effective, machine learning techniques.
machine learning Featured Proxy-Based Approach to Fair Machine Learning We explain the motivation behind our 2018 NeurIPS paper, "Hunting for Discriminatory Proxies in Linear Regression Models".
deep learning Featured Explainability in Neural Networks, Part 4: Path Methods for Feature Attribution This post will delve deeper into Path Integrated Gradient Methods for Feature Attribution in Neural Networks.
deep learning Featured Explainability in Neural Networks, Part 3: The Axioms of Attribution In this third post of the series on Explainability in Neural Networks, we present Axioms of Attribution, which are a set of desirable properties that any reasonable feature-attribution method should have.
deep learning Featured Explainability in Neural Networks, Part 2: Limitations of Simple Feature Attribution Methods We examine some simple, intuitive methods to explain the output of a neural network (based on perturbations and gradients), and see how they produce non-sensical results for non-linear functions.
deep learning Featured Explainability in Deep Neural Networks The wild success of Deep Neural Network (DNN) models in a variety of domains has created considerable excitement in the machine learning community. Despite this success, a deep understanding of why DNNs perform so well, and whether their performance is somehow brittle, has been lacking.
