Building systems with robustness, fairness, and privacy
CASC logo overlaid on abstract graphic of supercomputer racks
Center for Applied Scientific Computing

Assured Machine Learning: Robustness, Fairness, and Privacy

CASC is looking for creative team members at all stages of their careers. We invite you to browse the information below and apply to our open positions. Please contact us with any questions.

Scientific applications often have a broad range of real-world variations—data bias, noise, unknown transformations, adversarial corruptions, or other changes in distribution. Many of LLNL’s mission-critical applications are considered high regret, implying that faulty decisions can risk human safety or incur significant costs. As vulnerable ML systems are pervasively deployed, manipulation and misuse can have serious consequences.

A sustainable acceptance of ML requires evolving from an exploratory phase into development of assured ML systems that provide rigorous guarantees on robustness, fairness, and privacy. We’re using techniques from optimization, information theory, and statistical learning theory to achieve these properties, as well as designing tools to efficiently apply these techniques to large-scale computing systems.

Grid of images generated by AttGAN

Attribute-Guided Adversarial Training

More research accepted at AAAI 2021 enables DNNs to be robust against a wide range of naturally occurring perturbations.

NeurIPS logo

Two NeurIPS Acceptances

The 34th Conference on Neural Information Processing Systems features research on the reliability of DL for mission-critical applications.

part of illustration of the proposed UM-GNN showing a poisoned graph

Uncertainty-Matching GNNs

This paper, accepted at AAAI 2021, introduces the Uncertainty Matching Graph Neural Network aimed at improving the robustness of GNN models.

part of the Cifar-10C dataset shown as candlesticks

Feature Importance Estimation

In this paper accepted at AAAI 2021, a research team describes PRoFILE, a novel feature importance estimation method.

screen shot of part of the poster PDF

Advancing ML for Mission-Critical Applications

As our booth and poser at the 2020 LLNL Computing Virtual Expo explained, our work is a partnership between ML, HPC, and Lab programs.

charts showing comparisons of the marginals of posteriors

Calibration with NN Surrogates

Calibration of an agent-based epidemiological model (EpiCast) uses simulation ensembles for different U.S. metro areas. Peer review pending.

Jay presenting his research in front of an audience

Meet a Data Science Expert

Jay Thiagarajan’s research involves different types of large-scale, structured data that require the design of unique ML techniques.

collage of screen shots of NLPVis user interface

NLPVis Software Repository

NLPVis is designed to visualize the attention of neural network based natural language models.

Data Skeptic podcast logo

Data Skeptic Podcast

CASC researcher Jay Thiagarajan discusses reliable, interpretable predictive models for healthcare applications.