Robust Machine Learning

Feb 17, 2022

How can we make a machine learning model reliable? If generalization to unseen data is undoubtedly necessary, it is rarely sufficient. Models such as neural networks typically involve millions of operations to turn their input data into a prediction. This complexity permits to solve hard problems like computer vision and protein structure prediction. However, these complex models tend to exhibit unpredictable behaviours, such as the sensitivity to adversarial perturbations and spurious correlations . When models penetrate critical areas such as medicine, finance and the criminal justice system, this unpredictability is highly problematic.

Robust machine learning tackles this problem by providing tools to identify and fix the weaknesses of a model. This includes (but is not limited to) understanding the model’s failure modes, quantifying the model’s uncertainty and teaching a model to be robust with respect to noise/adversarial perturbations. With the development of Data-Centric AI, I believe that many new and exciting things can be done in this area.

Machine learning Deep learning Robustness Reliability Safety

Jonathan Crabbé

PhD Researcher

My research focuses on explainable artificial intelligence, representation learning and robust machine learning.

Posts

2 Papers Accepted at ICML 2022

Label-Free Explainability and Data-SUITE have been accepted at ICML 2022. These are two new methods that machine learning interpretability and robustness. I provide a quick summary of the methods along with the relevant links.

Jonathan Crabbé

Jun 8, 2022 1 min read Conferences

2 Papers Accepted at ICML 2022

Publications

Robust multimodal models have outlier features and encode more concepts

We investigate what makes multimodal models that show good robustness with respect to natural distribution shifts (e.g., zero-shot CLIP) different from models with lower robustness using interpretability.

Jonathan Crabbé, Pau Rodríguez, Vaishaal Shankar, Luca Zappella, Arno Blaas

Robust multimodal models have outlier features and encode more concepts

TRIAGE: Characterizing and auditing training data for improved regression

TRIAGE provides systematic data characterization for regression settings; compatible with a variety of regressors.

Nabeel Seedat, Jonathan Crabbé, Zhaozhi Qian, Mihaela van der Schaar

Evaluating the Robustness of Interpretability Methods through Explanation Invariance and Equivariance

We assess the robustness of various interpretability methods by measuring how their explanations change when applying symmetries of the model to the input features.

Jonathan Crabbé, Mihaela van der Schaar

Evaluating the Robustness of Interpretability Methods through Explanation Invariance and Equivariance

Joint Training of Deep Ensembles Fails Due to Learner Collusion

Training deep ensembles via a shared objective results in degenerate behavior.

Alan Jeffares, Tennison Liu, Jonathan Crabbé, Mihaela van der Schaar

Joint Training of Deep Ensembles Fails Due to Learner Collusion

Data-IQ: Characterizing subgroups with heterogeneous outcomes in tabular data

We introduce Data-IQ, a data-centric framework to identify ambiguous examples.

Nabeel Seedat, Jonathan Crabbé, Ioana Bica, Mihaela van der Schaar

Benchmarking Heterogeneous Treatment Effect Models through the Lens of Interpretability

We benchmark treatment effect models with interpretability tools.

Jonathan Crabbé, Alicia Curth, Ioana Bica, Mihaela van der Schaar

Data-SUITE: Data-centric identification of in-distribution incongruous examples

We introduce Data-SUITE, a data-centric framework to identify incongruous examples.

Nabeel Seedat, Jonathan Crabbé, Mihaela van der Schaar

Latent Density Models for Uncertainty Categorization

We introduce DAUX, an interpretability framework to interpret model uncertainty.

Hao Sun, Boris van Breugel, Jonathan Crabbé, Nabeel Seedat, Mihaela van der Schaar

Events

2 Spotlight Presentations at ICML 2022

Presentation of two accepted papers at the main conference.

Jul 17, 2022 — Jul 23, 2022 Baltimore Convention Center

2 Spotlight Presentations at ICML 2022