July 16, 2024
Congratulations to Distinguished Members of Technical Staff
Former DSI director Michael Goldman and DSI Council member Barry Chen were recently named as Distinguished Members of Technical Staff (DMTS) for their extraordinary scientific and technical contributions to the Lab. DMTS is the highest technical staff level achievable by a Livermore scientist or engineer.
Goldman founded the DSI in 2018 and was its first director until 2023, establishing a strong, forward-leaning vision and consistent funding stream. He has been a technical lead on several Global Security projects and programs in computer vision, imagery, machine learning (ML), large language models, and adversarial artificial intelligence (AI). He currently serves as the associate program leader for the Advanced Exploitation program, which brings ML to many of the Lab’s national security customers. Goldman consistently leverages his passion for workforce development to establish new pipelines through university partnerships, build mentoring and upskilling programs for existing staff, and create opportunities to strengthen the Lab’s data science community. He received his M.S. from UC Davis in electrical and computer engineering.
Chen is an ML researcher with over 19 years of experience in developing and applying novel algorithms to a wide variety of projects and applications at LLNL predominantly in the Global Security Directorate. With expertise in neural networks, random forests, and probabilistic graphical models, Chen has helped advance AI to enhance threat detection, prediction, and analysis capabilities. He currently leads several research teams developing new neural network learning algorithms with scientific and security applications. Chen holds a PhD from UC Berkeley and has been a member of the DSI Council since its inception.
“Mike and Barry are changemakers and leaders in the data science field, both in research and in their engagement with the broader community. Their research and contributions over the years have significantly elevated data science, internally and externally, and the DSI would not exist without their hard work and dedication to our field,” says DSI deputy director Cindy Gonzales.
Video: The Surprising Places You’ll Find Machine Learning
LLNL data scientists are applying ML to real-world applications on multiple scales. A new DSI-funded video highlights innovative research at the nanoscale (developing better water treatment methods by predicting the behavior of water molecules under the extremely confined conditions of nanotubes); mesoscale (determining the likelihood and location of a dangerous wildfire-causing phenomenon called arcing); and macroscale (simulating methods for increasing efficiency for the removal of carbon from the atmosphere). Watch as researchers Anh Pham, Indra Chakraborty, and Youngsoo Choi explain why problems like water filtration, wildfires, and carbon capture are becoming more solvable thanks to groundbreaking data science methodologies on some of the world’s fastest computers.
AI Roundtable with Silicon Valley Leadership Group
DSI deputy director Cindy Gonzales and AI Innovation Incubator director Brian Spears recently participated in an AI roundtable discussion with California state senator Anna Caballero and the Silicon Valley Leadership Group (SVLG). Among the discussion topics was the need to educate legislators and their staffers on AI/ML concepts and technologies, which would help inform policy decisions. Another need the group recognized was education/retraining programs for California residents with nontraditional backgrounds, which could present an opportunity for the DSI’s Data Science Challenge (DSC) program.
“The DSC is a perfect fit for this model. As we plan to expand our training programs next fiscal year, we could feasibly offer DSC to someone with little to no background in the field and help them reach to a level of comfortability in working with data and using well-known tools,” Gonzales notes. “The potential is endless, and partnering with local government officials to understand the needs and how we can help is a part of our larger collective responsibility.”
Finding the Sweet Spot in AI Model Optimization
It’s almost an understatement to say that expectations are high for AI systems. For example, AI models need to accurately detect anomalous data compared to the training dataset, and they need to generalize to previously unseen (out-of-distribution, or OOD) data. Optimizing models for one of these goals often comes at the expense of the other. However, a Livermore-led research team has figured out how to balance this tradeoff by leveraging model anchoring.
The team’s paper, “The Double-Edged Sword of AI Safety: Balancing Anomaly Detection and OOD Generalization via Model Anchoring,” was accepted to the 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). The authors are Vivek Narayanaswamy and Jayaraman Thiagarajan alongside former LLNL computer scientist Rushil Anirudh, now at Amazon.
Anchoring is a methodology for training deep neural networks that re-parameterizes the input into anchor–residual pairs. Here the anchor is drawn at random from the training dataset itself and the residual is the difference between the input and the anchor. The team further modified this method with perturbed anchoring (PA) and residual regularization (RR) techniques. PA corrupts the anchor without modifying the residual, while RR masks the anchor after computing the residual. Controlling individual components of the pairs enables finer control of the model’s dependency on data distributions.
As shown in the image at left, the team’s two-pronged anchoring method improves both anomaly detection and OOD generalization when PA and RR are used judiciously. The x-axis shows the ratio of how many times each method is invoked during training, while the y-axis tracks accuracy (left) and 100-FPR (right) scores against standard benchmarking datasets. Using only RR (left-hand pink area) does not provide the level of optimizations as with using both PA and RR (light green area). As the PA:RR ratio climbs, however, detection accuracy dramatically drops off and only generalization accuracy continues to improve (right-hand pink area).
“We found that we can selectively improve anomaly detection and generalization via a novel anchored training mechanism without exposing models to additional outlier data or incurring additional computational cost,” explains Narayanaswamy. “There’s a nontrivial relationship between these objectives, which are crucial to AI interpretability and trustworthiness.”
Data Science Consultants Spread the Word
Approximately 60 LLNL employees turned out for the Data Science Institute Consulting Service’s (DSICS) “mini road shows”—events designed to raise awareness about DSI-funded consulting services, solicit projects, and recruit new consultants. The road shows described the consulting process from different perspectives, including how projects can request help and how individuals can become consultants. To cap the presentations, consultant Mike Boyle shared his story of quickly developing a database of molecular weights to enable an analysis by the Forensic Science Center (see story in newsletter volume 35).
Giselle Fernandez, staff scientist and a DSICS deputy director, notes, “Our recent road shows promoting our consulting services were successful, with enthusiastic attendance that exceeded our expectations and generated strong interest from potential consultants and consultees. A highlight was Mike’s testimony, whose exceptional contributions as a consultant earned him a Global Security Bronze Award, exemplifying the powerful collaboration between domain science and data science.”
Attendees who expressed interest in becoming consultants have already had the opportunity to do so. Tyler Alcorn, a health physicist with a background and continued interest in data science, recently joined a consulting project on analyzing scintillator data. “I’m thankful for the opportunity to work with DSICS as it allows me to collaborate with other programs and scientists, and to be a part of interesting projects outside my current job scope. DSICS also gives me a chance to flex my data science skillset and to be mentored by more senior data scientists. It's been a very positive experience!” he says.
The road show will likely become a recurring event to help identify new LLNL projects in need of data science expertise and to inspire future consultants. DSICS leaders are also planning a follow-up event to train new consultants.
Drug Design Milestone Relies on AI and HPC
In a substantial milestone for supercomputing-aided drug design, LLNL and BridgeBio Oncology Therapeutics announced that clinical trials have begun for a first-in-class medication that targets specific genetic mutations implicated in many types of cancer. The development of the new drug—BBO-8520—is the result of collaboration among LLNL, BridgeBio, and the National Cancer Institute (NCI)’s RAS Initiative at the Frederick National Laboratory for Cancer Research (FNL). In a first for a DOE national laboratory, the drug was discovered through DOE’s leadership in high-performance computing (HPC) for mission applications, combined with an LLNL-developed platform integrating AI and traditional physics-based drug discovery, and effective partnership with the FNL and NCI.
The drug candidate has shown promise in laboratory testing for inhibiting mutations of KRAS proteins linked to about 30% of all cancers—targets long considered “undruggable” by cancer researchers. The achievement provides hope for broad impact on cancer patients whose tumors harbor susceptible KRAS mutations. This indicates that a computational/AI drug design approach could unlock new insights into the disease and the future of cancer treatment.
In addition to advancing cancer research, LLNL representatives said the milestone is validation that integrating supercomputing with AI- and physics-based computational platforms has the potential to further accelerate small-molecule drug discovery and equip DOE, the National Nuclear Security Administration, and LLNL with the ability to quickly and routinely develop medical countermeasures for disease or future pandemics, aligning with broader mission focus areas in biosecurity, bioresilience, and national security.
Workshop Spotlights Signal and Image Sciences
Nearly 150 members of the signal and image sciences community recently came together to discuss the latest advances in the field and connect with colleagues, friends, and potential collaborators at the 28th annual Center for Advanced Signal and Image Sciences (CASIS) workshop. The event featured more than 50 technical contributions across six workshop tracks and a parallel tutorials session, including 40 talks and 23 posters that helped encourage discussions. This year’s topics included remote and noninvasive sensing, nondestructive evaluation, signal and image sciences at the National Ignition Facility (NIF), AI/ML, quantum sensing, and quantum computing and energy applications. Signal and image sciences enable efficient and accurate processing, generation, analysis, and interpretation of signals and images in fields such as telecommunications, medical imaging, computer vision, and more. At the Lab, they are the backbone of NIF diagnostics, nondestructive evaluation and characterization, advanced sensing, AI/ML, and various other critical mission roles. CASIS and the DSI recently co-sponsored an AI safety workshop and seminar (see next story).
Seminar Video: How Could We Design Aligned and Provably Safe AI?
On April 19, Dr. Yoshua Bengio presented “How Could We Design Aligned and Provably Safe AI?” His seminar was co-sponsored by the DSI and CASIS. A Turning Award winner, Bengio is recognized as one of the world’s leading AI experts, known for his pioneering work in deep learning. He is a full professor at the University of Montreal, and the founder and scientific director of the Mila–Quebec AI Institute. In 2022, Bengio became the most-cited computer scientist in the world.
Bengio discussed his AI research program based on risk evaluation, Bayesian priors, and uncertainty, as well as how amortized inference with large neural networks could be made to estimate the required quantities. A video of his talk is now available on YouTube. Seminar speakers’ biographies and abstracts are available on the seminar series web page, and many recordings are posted to the YouTube playlist. To become or recommend a speaker for a future seminar, or to request a WebEx link for an upcoming seminar if you’re outside LLNL, contact DSI-Seminars [at] llnl.gov (DSI-Seminars[at]llnl[dot]gov).
Recent Research
- Adversarial Robustness Limits via Scaling-Law and Human-Alignment Studies, 2024 International Conference on Machine Learning (ICML 2024) – Brian Bartoldson, James Diffenderfer, Konstantinos Parasyris, and Bhavya Kailkhura
- Decoding Compressed Trust: Scrutinizing the Trustworthiness of Efficient LLMs Under Compression, ICML 2024 – Diffenderfer, Bartoldson, and Kailkhura with colleagues from multiple universities
- On the Fly Neural Style Smoothing for Risk-Averse Domain Generalization, 2024 IEEE Winter Conference on Applications of Computer Vision (WACV 2024) – Kailkhura with colleagues from Tulane University
- On the Use of Anchoring for Training Vision Models, preprint – Vivek Narayanaswamy, Kowshik Thopalli, Yamen Mubarka, Wesam Sakla, and Jayaraman Thiagarajan with a colleague from Amazon
- PAGER: Accurate Failure Characterization in Deep Regression Models, ICML 2024 – Thiagarajan and Narayanaswamy with colleagues from Amazon and the University of Michigan
- Q-Hitter: A Better Token Oracle for Efficient LLM Inference via Sparse-Quantized KV Cache, 2024 Conference on Machine Learning and Systems (MLSys) – Kailkhura with colleagues from several universities
- Towards Anatomy Education with Generative AI-Based Virtual Assistants in Immersive Virtual Reality Environments, 2024 IEEE International Conference on Artificial Intelligence and eXtended and Virtual Reality (AIxVR 2024) – Vuthea Chheang with colleagues from the University of Delaware
- Towards Generalizable and Interpretable Motion Prediction: A Deep Variational Bayes Approach, Proceedings of Machine Learning Research – Yeping Hu with colleagues from Purdue University and UC Berkeley
- Transformers Can Do Arithmetic with the Right Embeddings, preprint – Bartoldson and Kailkhura with colleagues from three institutions (image at left: state-of-the-art embeddings compared to the team’s solution, with greater accuracy shown in blue)
- TrustLLM: Trustworthiness in Large Language Models, ICML 2024 – Kailkhura with colleagues from multiple institutions and universities
Award-Winning Data Science Solutions
LLNL’s HPC and data science capabilities play a significant role in international science research and innovation, and Lab researchers have won 10 R&D 100 Awards in the Software–Services category in the past decade. The latest issue of Science & Technology Review features several award-winning projects, including ZFP and CANDLE.
ZFP introduces a new method of compressing large datasets while maintaining high-speed, on-demand access to the compressed data for both reading and writing applications—a capability not found in any other compressor. Researchers can continue to work with the data in real time while it remains compressed, whether they are analyzing it or producing visualizations. ZFP is downloaded more than 1.5 million times per year by users from across the DOE and other government and nongovernment agencies, and its scientific applications include geographic information systems, climate science, seismology, and tornado simulations, among others.
An early adopter of using ML for scientific applications, the Cancer Distributed Learning Environment (CANDLE) provides ML capabilities for applications related to cancer research. In particular, CANDLE enables capabilities for extracting key information and finding relationships within large, disconnected datasets to help solve cancer-specific drug challenges. CANDLE is a collaboration among Lawrence Livermore, Los Alamos, Oak Ridge, and Argonne national laboratories; the Frederick National Laboratory for Cancer Research; the National Institutes of Health; and the National Cancer Institute.
Watch the WiDS Livermore Video Playlist
If you missed our Women in Data Science (WiDS) datathon in February or the WiDS Livermore conference in March, videos of the technical talks, panel discussions, and a dataset tutorial are posted to the playlist on YouTube. There are 10 new videos in the playlist, which also includes highlights from our 2022 and 2023 events.
The global WiDS Conference aims to inspire and educate data scientists worldwide, regardless of gender, and to support women in the field. WiDS Livermore is independently organized by LLNL to be part of the mission to increase participation of women in data science and to feature outstanding women doing outstanding work. Learn more on the WiDS Livermore web page.
Meet an LLNL Data Scientist
Aneesha Devulapally is a data scientist in the Global Security Computing Applications Division. She is particularly interested in the interdisciplinary field of bioinformatics and applications of ML in systems biology. As a part of LLNL’s GUIDE program, she develops ML frameworks and pipelines and performs data analysis on antibody–antigen complexes. “Data science brings subject matter experts from various domains together to collaborate in solving complex problems, especially at Livermore where we’re working towards solving problems for the betterment of humankind,” Devulapally says. She joined LLNL after earning her master’s in computer science with a specialization in data science from the University of Texas at Dallas after earning undergraduate and graduate degrees at IIIT Bangalore in India. Devulapally recently served on the organizing committee for Livermore’s 2024 Women in Data Science event. “The enthusiasm of the participants and their interest in learning data science concepts made the experience incredibly rewarding,” she says.