We are currently witnessing one of the most rapid transformations of our technological landscape—AI is ubiquitous, appearing on our phones, our computers, in the news, and in our conversations with family and friends.

The incredible rise of advanced AI models like ChatGPT—which receives nearly 2 billion web visits monthly—points to just one tangible example of how AI is reshaping our world. However, such advanced AI models would not be possible without machine learning (ML), the intellectual engine that enables these systems to generate outputs through pattern recognition and the analysis of enormous datasets. 

At the forefront of this changing landscape are researchers from Pacific Northwest National Laboratory (PNNL) who are committed to advancing technology such as AI and ML to tackle our most challenging and pressing problems. Recently, PNNL researchers showcased their expertise and research at the prestigious International Conference on Machine Learning (ICML) in Vancouver, BC, on July 13–19.

Disruptive tech requires significant consideration  

Beyond the typical ethical and economic considerations that dominate most conversations surrounding AI and ML are considerations surrounding use cases and opportunities in application areas such as speech recognition, robotics, and computational biology. Events like ICML serve as crucial platforms for exploring such possibilities.

The ICML, distinguished as the oldest and fastest growing conference of its kind, has established itself as a central hub for experts to gather annually and discuss the latest research in ML and its related domains: AI, statistics, and data science.

Among the attendees, spanning a wide range of backgrounds, were several data scientists, computer scientists, and mathematicians from PNNL who contributed to the event with featured presentations and posters.

AI to advance mathematics

Leading the featured presentation “Machine Learning meets Algebraic Combinatorics: A Suite of Datasets Capturing Research-level Conjecturing Ability in Pure Mathematics” was AI researcher and mathematician Henry Kvinge, who has a deep-seated interest in the intersection of AI and mathematics, both in terms of the use of mathematics to better understand AI systems as well as the use of AI to advance mathematics.

Bridging the gap 

Kvinge’s presentation was supported by collaborative efforts from former PNNL intern Herman Chau, intern Jesse He, and data scientists at PNNL: Helen Jenne, Davis Brown, and Mark Raugas. Together, the team constructed a collection of datasets—the Algebraic Combinatorics Dataset Repository—representing the process of research-level conjecture generation. Each dataset consists of millions of examples of polynomials, graphs, or lattice paths that represent open problems in combinatorics and can be used to test the ability of AI systems to provide insights to mathematicians.

“These datasets were designed to highlight an aspect of the mathematician’s workflow—data exploration and conjecture generation—that has mostly been neglected by AI for math-related benchmarks,” Kvinge explained. “We ultimately see AI as a tool that is analogous to a telescope, expanding the scope and scale of information from which a mathematician can extract insight.”

Increasing what is possible  

The team’s poster “Machines and Mathematical Mutations: Using GNNs to Characterize Quiver Mutation Classes” was also accepted into the event. The research, presented by He, sought to push the boundaries of computational mathematics by leveraging ML and graph neural networks to explore the intricate patterns of quiver mutations.

Speaking on the importance of their research, He added, “Our work presents a really interesting case study of how ML can distill a huge number of examples—more than a human could look at individually—into simple rules that humans can then prove.”

Greater accuracy and sustainability  

PNNL intern Chuan Liu presented the final PNNL featured poster, “An Expressive and Self-Adaptive Dynamical System for Efficient Function Learning.” Working in collaboration with PNNL computer scientist Ang Li, they developed an expressive and self-adaptive dynamical system that is not only more sustainable than traditional methods but also delivers greater accuracy.

Speaking on the inspiration and capabilities of the system, Liu said, “Drawing inspiration from biological systems, our approach integrates hierarchical architectures with heterogeneous dynamics to enhance the system’s expressivity and employs intrinsic electrical signals for on-device training, thereby enabling accurate and efficient learning of various functions.”

Continued advancements 

PNNL’s contributions to ICML—from advancing pure mathematics to developing more sustainable computational systems—represent just a glimpse of the Lab’s innovative AI and ML research tackling fundamental scientific challenges.

To learn more about PNNL’s AI and ML research, visit our dedicated data science and computing page. For additional information about ICML and its featured presentations, visit the conference website.