By: Eli Chien
December 19, 2024

Vancouver, Canada – The Thirty-Eighth Annual Conference on Neural Information Processing Systems (NeurIPS 2024) was held in Vancouver from December 10 to 15, 2024. NeurIPS is one of the premier machine learning and AI conferences, gathering outstanding researchers from around the world. A3D3 members contributed seven posters, including one spotlight poster and one benchmark poster, and three workshop papers. In addition, A3D3 member Dr. Thea Klaeboe Aarrestad from ETH Zürich was an invited speaker at the WiML (Women in Machine Learning) workshop and the ML4PS (Machine Learning and the Physical Sciences) workshop, presenting “Pushing the limits of real-time ML: Nanosecond inference for Physics Discovery at the Large Hadron Collider.” We are excited to see that A3D3 members are an important part of the machine learning community!

A3D3 members gather at NeurIPS 2024.

Invited Talks

Dr. Thea Klaeboe Aarrestad gives invited talks for the WiML workshop (left) and the ML4PS workshop (right).

Poster Presentations

Mingfei Chen (left) and Dr. Eli Shlizerman (right) from the University of Washington presented “AV-Cloud: Spatial Audio Rendering Through Audio-Visual Cloud Splatting.” They propose a novel approach, AV-Cloud, for rendering high-quality spatial audio in 3D scenes that is in synchrony with the visual stream but does not rely on nor is explicitly conditioned on the visual rendering. [paper link, project link]

Dr. Eli Chien from the Georgia Institute of Technology presented a spotlight poster titled “Langevin Unlearning: A New Perspective of Noisy Gradient Descent for Machine Unlearning.” They propose Langevin unlearning which not only connects DP with unlearning, but also provides unlearning privacy guarantees for PNGD with Langevin dynamic analysis. [paper link]

Eli Chien also presented a poster titled “Certified Machine Unlearning via Noisy Stochastic Gradient Descent.” They propose to leverage stochastic noisy gradient descent for unlearning and establish its first approximate unlearning guarantee under the convexity assumption. [paper link]

Jingru (Jessica) Jia from the University of Illinois at Urbana-Champaign presented “Decision-Making Behavior Evaluation Framework for LLMs under Uncertain Context.” This paper quantitatively assesses the decision-making behaviors of large language models (LLMs), demonstrating how these models mirror human risk behaviors and exhibit significant variations when embedded with socio-demographic characteristics. [paper link]

Shikun Liu from the Georgia Institute of Technology presented “GeSS: Benchmarking Geometric Deep Learning under Scientific Applications with Distribution Shifts.” They propose GeSS, a comprehensive benchmark designed for evaluating the performance of geometric deep learning (GDL) models in scientific scenarios with various distribution shifts, spanning scientific domains from particle physics and materials science to biochemistry. [paper link]

Yingbing Huang from the University of Illinois Urbana-Champaign presented “SnapKV: LLM Knows What You are Looking for Before Generation.” They propose SnapKV, a fine-tuning-free method to reduce the Key-Value (KV) cache size in Large Language Models (LLMs) by clustering important positions for each attention head. [paper link]

Xiulong Liu from the University of Washington presented “Tell What You Hear From What You See – Video to Audio Generation Through Text.” It is a novel multi-modal generation framework for text guided video-to-audio generation and video-to-audio captioning. [paper link]

Workshop Presentations

Zihan Zhao from the University of California, San Diego presented “Learning Symmetry-Independent Jet Representations via Jet-Based Joint Embedding Predictive Architecture” at ML4PS workshop. They propose an approach to learning augmentation-independent jet representations using a jet-based joint embedding predictive architecture (J-JEPA). [paper link, project link]

Trung Le from the University of Washington presented “NetFormer: An interpretable model for recovering identity and structure in neural population dynamics” at the NeuroAI workshop. They introduce NetFormer, an interpretable dynamical model to capture complicated neuronal population dynamics and recover nonstationary structures. [paper link]

Andrew “AJ” Wildridge from Purdue University presented “Bumblebee: Foundation Model for Particle Physics Discovery” at ML4PS workshop. Bumblebee is a transformer-based model for particle physics that improves top quark reconstruction by 10-20% and effectively generalizes to downstream tasks of classifying undiscovered particles with excellent performance. [paper link, project link]

Patrick Odagiu from ETH Zurich presented “Knowledge Distillation for Teaching Symmetry Invariances” at SciforDL Workshop. They show that knowledge distillation is just as good as data augmentation for learning a specific symmetry invariance in your data set. [paper link]

By: Cymberly Tsai
December 19, 2024

About the author: Cymberly Tsai is a high school senior at Council Rock High School North working with Prof. Joshua Agar at Drexel University. Cymberly presented a poster titled “Foundation Model for Real-Time Model Selection and Fitting” at the A3D3 All-Hands Meeting at Purdue University.

Participants of the 2024 A3D3 All-Hands Meeting gather for a group photo.

The A3D3 Institute hosted its annual All-Hands Meeting at Purdue University on October 14, 2024. The conference was a constellation of ideas as participants presented bright projects, which contributed to a sky of innovation, and guided others at various stages of education. As a student at Council Rock High School North, I had the opportunity to attend this conference with the Drexel Multifunctional Materials and Machine Learning lab with professor Josh Agar, after working with the lab the past summer. Taking place at the intersection of artificial intelligence and scientific research, the conference included attendees from a vast range of universities and locations, offering both in-person and virtual participation.

The conference emphasized accelerating machine learning for applications in a variety of fields. Located at Purdue’s Convergence Center, the meeting began with talks from several principal investigators, including Dr. Song Han‘s talk introducing efficient multi-modal large language model (LLM) innovations and Dr. Pan Li’s work on Graph Machine Learning under distribution shifts. These in-depth presentations offered insight into specific innovations, inspiring further research from attendees. Highlighting many varied applications of machine learning, trainees presented overviews of fields with high potential for fast machine learning, such as multi-messenger astronomy, neuroscience, high-energy physics, and hardware/algorithm co-development. The trainee-led talks covered such cutting-edge topics as the challenges of processing massive data rates in high-energy physics research and real-time brain modeling in primates with machine learning accelerated by Field-Programmable Gate Arrays (FPGAs). As one of the five National Science Foundation’s (NSF) Harnessing the Data Revolution (HDR) institutes, A3D3 also announced the NSF HDR ML Anomaly Detection Challenge at the conference, which leverages datasets in multiple fields to encourage data-driven discovery.

The All-Hands Meeting was particularly beneficial for those early in their research careers, welcoming them into the academic community. The conference offered a poster session, providing a chance to engage with ideas in a low-stress setting. Beyond research-specific content, the conference included a panel discussion on how to apply to faculty, postdoc, and industry jobs, led by Dr. Zhijian Liu, an NVIDIA research scientist and incoming assistant professor at UCSD, Dr. Julia Gonski, a Panofsky Fellow at SLAC/Stanford, and Dr. Dylan Rankin, an assistant professor at the University of Pennsylvania. Postbaccalaureate fellows of A3D3 also described their experience, giving insight into another avenue of research. A highlight of the All-Hands Meeting was the trainee-led townhall, which gave trainees and early-career researchers a chance to discuss ideas such as potential presentation opportunities and future collaborations.

The A3D3 conference experience expanded my understanding of machine learning and the world of research, sparking a greater curiosity and motivation to pursue scientific advancements with artificial intelligence. Following the All-Hands Meeting, the Fast Machine Learning for Science Conference conveniently took place at Purdue from October 15 through 18, allowing A3D3 attendees to dive deeper into machine learning advancements in research. Attending the conference was an invaluable experience for me to engage in the world of research and witness innovations in machine learning that I hope to contribute to in the future.

By: Miles Cochran-Branson
July 26, 2024

Following the three-day US-ATLAS conference held at the University of Washington (UW) Seattle campus, students met for a brief tutorial on tools available for US-ATLAS members. After introductions in computing resources, Yuan-Tang Chou—a postdoc at UW and member of the A3D3 team—gave a presentation on GPU resources and utilization with a focus on new applications to particle-physics workflows. 

Chou gives a brief presentation on applications of the NVIDIA Triton server for deploying models as-a-service.

The presentation focused on deploying models on accelerators such as GPUs, as-a-service (aaS) using the NVIDIA Triton Inference Server. Chou discussed the merits of heterogeneous computing—the most straightforward way to deploy algorithms where the CPU and GPU both are connected on a single node. He noted that for many physics processes such as Graph-Neural-Network (GNN) based tracking, flavor tagging, and detector simulation, heterogeneous computing could be “inefficient and very expensive to scale.” Hence, offloading expensive tasks to a GPU server could streamline the deployment of large models important for ATLAS physics. 

Sample architecture of as-a-service model deployment from CPU-only client nodes or CPU / GPU client nodes to a GPU server.

After motivating why deploying models as-a-service could be beneficial in physics analysis, Chou gave a brief demo on deploying GNN-tracking aaS. This was followed by a hands-on tutorial deploying the resnet50 image recognition deep neural network as-a-service on computing resources at CERN. The tutorial material focused on building the proper model repository structure and configuration for image detection on a GPU server. Students set-up a work environment, deployed a backend on the server, and sent an image to the server to be classified. 

Students work on deploying a backend on a server and sending images to the server to be classified. 

By the end of the tutorial, students had successfully deployed a backend and received image classifications back from the image recognition model. Interested students were connected with experts currently working in ongoing development of aaS tools in algorithm development of ATLAS algorithms.

Tutorial Resources: https://hrzhao76.github.io/AthenaTriton/triton-lxplusGPU.htmlDeveloped by Yuan-Tang Chou, Miles Cochran-Branson (UW), Xiangyang Ju (LBNL), and Haoran Zhao (UW)

Written by Miles Cochran-Branson, PhD student at University of Washington