Soumyendu Sarkar’s AI research team at Hewlett Packard Labs will present 10 papers at this week’s prestigious Conference on Neural Information Processing Systems (NeurIPS) in Vancouver. This annual event is known widely as one of the world's top conferences for AI/ML and data science.
Three of the papers focus on how artificial intelligence (AI) can advance sustainability of data centers. Four papers address large language models (LLMs) and foundational models spanning different areas from trustworthiness, multi-modal video-language models, and application of LLMs for analytics and code generation for nuclear fusion optimization. Two papers involve collaborations with University of Rochester on a U.S. Department of Energy-funded project to advance nuclear fusion research with AI with a promise to unlimited clean energy. The tenth paper covers collaborative work with Carnegie Clean Energy to advance wave energy production.
“This is a huge presence for our team at NeurIPS,” said Sarkar, a senior distinguished technologist at Hewlett Packard Labs. “Our four papers on Generative AI and large language models cut across both trust in these technologies and how the power of these technologies can be applied to real world use cases to derive value from AI investments. We are making a difference not only by delivering advanced technologies but also by applying cutting-edge AI research for the good of mankind, all the way from sustainable cloud computing to clean energy advancement with nuclear fusion and wave energy. I want to thank my team for making this possible with their world class research talent. They are driven by HPE’s vision to make world a better place.”
NeurIPS brings together some of the foremost thought leaders in AI and data science. It provides a platform for researchers, scientists, engineers, and industry professionals to share their latest findings, exchange ideas and collaborate on cutting-edge research.
Exploring sustainability for data centers and beyond
The selected main conference paper, titled “SustainDC: Benchmarking for Sustainable Data Center Control,” advances the frontier of holistic real-time data center management for sustainability using reinforcement learning, a key technology behind self-driving cars.
“What’s significant is this work makes customizable data center modelling available with an AI framework via open source, so other researchers and other collaborators in the ecosystem can easily build digital twins of their own data centers and test their innovations,” Sarkar said. “This also bridges the gap of bringing cutting-edge AI research to cloud computing sustainability, by democratizing access to data center models and benchmarks, enabling AI research communities, who are not domain experts, to contribute effectively.”
This work builds on the team’s previous paper “Real-Time Carbon Footprint Minimization in Sustainable Data Centers with Reinforcement Learning,” which received the Best ML Innovation paper award at with 2023 NeurIPS workshop on Tackling Climate Change with Machine Learning. The path forward to this research is the control optimization of a geographically distributed data center cluster.
This research will also be presented in the NeurIPS 2024 flagship Climate Change workshop in a paper titled “Carbon-Aware Spatio-Temporal Workload Distribution in Cloud Data Center Clusters Using Reinforcement Learning.” The Labs team members will further showcase their AI innovations on liquid cooling optimization in another paper at this same workshop titled “Enhancing Sustainability in Liquid-Cooled Data Centers with Reinforcement Learning Control.”
Putting the spotlight on Generative AI
Labs team members are presenting at the Red Teaming GenAI workshop, showcasing their innovation on automating red teaming for LLMs used in ChatGPT. Red teaming of LLMs involves systematically testing and probing these models to identify vulnerabilities, biases, and potential harms, ensuring safer and more reliable deployment. The paper is titled “iART – Imitation Guided Automated Red Teaming.”
“This work pioneers automated red teaming for LLMs to deliver diverse, high-quality testing with enhanced computational efficiency, ensuring safer AI deployments,” Sarkar said. Related research on automated red teaming will also be presented on Dec. 14 at the Workshop on New Frontiers in Adversarial Machine Learning, and Dec. 15 at the Safe Generative AI Workshop.
At the Adaptive Foundation Models workshop, the team will present how large language models can be applied to solve complex reasoning problems in a paper titled “Informed Tree of Thought: Cost-efficient Problem Solving with Large Language Models.”
“This paper presents the application of a reasoning framework where the model explores multiple reasoning paths and evaluates them hierarchically, enabling more deliberate and structured problem-solving,” Sarkar said. “Techniques like this can enable root cause analysis of data center operations from the system logs.”
The Labs team will also present other works related to GenAI, including automated code generation by LLMs to solve optimization problems in nuclear fusion and reinforcement learning-based agents to evaluate multi-modal vision-language models.
Advancing nuclear fusion and clean energy with AI research and collaboration
The Labs team will present two papers on advancing nuclear fusion with AI research at the workshop on Machine Learning for Physical Sciences and Climate Change. This is a U.S. Department of Energy-funded collaborative work with the physicists at the University of Rochester. Inertial confinement fusion (ICF) experiments are complex and exceptionally costly because of the extreme conditions needed to achieve nuclear fusion. This research aims to overcome the challenge of minimal experimental opportunity with advanced sample-efficient AI techniques that leverage existing knowledge to drastically reduce the experiments required for optimizing energy yield in fusion, which has promise for limitless clean energy.
The second paper presents a technique to combine machine learning with human expertise, in an AI-enabled optimization for fusion, where the model explains its recommendation and puts the scientist in the decision-making loop.
Another paper focuses on Hewlett Packard Labs’ ongoing work with Carnegie Clean Energy (CCE). Labs and CCE have teamed up to develop a self-learning wave energy converter using deep reinforcement learning technology. This research enhances the performance of reinforcement learning controllers for wave energy converters, by predicting incoming waves in real time. CCE had received a Phase 3 EuroWave contract to build and operate a wave energy prototype at a European wave energy test site in Spain and plans to extend AI research collaboration with HPE.
The schedule
The NeurIPS conference runs from Dec. 10 through Dec. 15. Sarkar’s AI team will present the following works:
December 11:
“SustainDC: Benchmarking for Sustainable Data Center Control,” by Avisek Naug, Antonio Guillen, Ricardo Luna Gutierrez, Vineet Gundecha, Cullen Bash, Sahand Ghorbanpour, Sajad Mousavi, Ashwin Ramesh Babu, Dejan Markovikj, Lekhapriya Dheeraj Kashyap, Desik Rengarajan, Soumyendu Sarkar.
SustainDC is an innovative optimization platform leveraging real-time multi-agent reinforcement learning to enhance data center sustainability. This pluggable, expandable digital twin democratizes machine learning research for sustainable operations. It contributes to the Frontier supercomputer's digital twin and the ExaDigiT consortium led by Oak Ridge National Labs. The foundational research won the Best Paper Award at last year’s NeurIPS workshop on Tackling Climate Change with Machine Learning.
- LINK: NeurIPS Poster SustainDC: Benchmarking for Sustainable Data Center Control
- VIDEO: https://youtu.be/KuvDDs-rjPU
December 14:
“Informed Tree of Thought: Cost-Efficient Problem Solving with Large Language Models,” by Sajad Mousavi, Desik Rengarajan, Ashwin Ramesh Babu, Sahand Ghorbanpour, Vineet Gundecha, Avisek Naug, Soumyendu Sarkar.
iToT, the Informed Tree of Thought, enhances large language models for dynamic, cost-efficient decision-making in multi-cloud IT operations with full-stack observability, laying the groundwork for HPE’s next-generation private cloud analytics.
“Coordinated Robustness Evaluation Framework for Vision Language Models,” by Ashwin Ramesh Babu, Sajad Mousavi, Desik Rengarajan, Vineet Gundecha, Sahand Ghorbanpour, Avisek Naug, Antonio Guillen-Perez, Ricardo Luna Gutierrez, Soumyendu Sarkar.
This work improves vision-language model robustness, ensuring reliable, trusted AI performance in sensitive applications like healthcare and autonomous systems.
December 15:
“iART – Imitation Guided Automated Red Teaming,” by Sajad Mousavi, Desik Rengarajan, Ashwin Ramesh Babu, Vineet Gundecha, Avisek Naug, Sahand Ghorbanpour, Ricardo Luna Gutierrez, Antonio Guillen, Paolo Faraboschi, Soumyendu Sarkar
This work is advancing AI safety for large language models with iART, pioneering automated red teaming for large language models to deliver diverse, high-quality testing and enhanced computational efficiency, ensuring safer AI deployments.
“Carbon-Aware Spatio-Temporal Workload Distribution in Cloud Data Center Clusters Using Reinforcement Learning,” by Soumyendu Sarkar, Antonio Guillen, Vineet Gundecha, Avisek Naug, Ricardo Luna Gutierrez, Sajad Mousavi, Paolo Faraboschi, Cullen Bash. This research leverages hierarchical reinforcement learning techniques to dynamically optimize workload distribution across cloud data center clusters while simultaneously running SustainDC like control optimizations on individual data centers, reducing carbon emissions and other sustainability improvements.
“Enhancing Sustainability in Liquid-Cooled Data Centers with Reinforcement Learning Control,” by Avisek Naug, Antonio Guillen, Vineet Gundecha, Ricardo Luna Gutierrez, Ashwin Ramesh Babu, Sajad Mousavi Paolo Faraboschi, Cullen Bash, Soumyendu Sarkar. This research introduces intelligent liquid cooling technology powered by reinforcement learning to optimize energy use while maintaining ideal server temperatures.
“Explainable Meta Bayesian Optimization with Human Feedback for Scientific Applications like Fusion Energy,” by Ricardo Luna Gutierrez, Sahand Ghorbanpour, Vineet Gundecha, Rahman Ejaz, Varchas Gopalaswamy, Riccardo Betti, Avisek Naug, Desik Rengarajan, Ashwin Ramesh Babu, Paolo Faraboschi, Soumyendu Sarkar. This is a U.S. DoE-funded collaborative work with the physicists at the University of Rochester. This research introduces a sample-efficient method for optimizing complex scientific processes like nuclear fusion, combining machine learning with human expertise to ensure experts remain involved in decision-making while providing explanatory insights. It integrates knowledge from multi-fidelity simulations, experiments, and scientific input.
“Enhancing Reinforcement Learning-Based Control of Wave Energy Converters Using Predictive Wave Modeling”, by Vineet Gundecha, Arie Paap, Mathieu Cocho, Sahand Ghorbanpour, Alexandre Pichard, Ashwin Ramesh Babu, Soumyendu Sarkar. This is a collaborative work with Carnegie Clean Energy. This research enhances the performance of reinforcement learning controllers for wave energy converters by predicting incoming waves in real time.
“Meta-Learned Bayesian Optimization for Energy Yield in Inertial Confinement Fusion,” by Vineet Gundecha, Ricardo Luna Gutierrez, Sahand Ghorbanpour, Desik Rengarajan, Rahman Ejaz, Varchas Gopalaswamy, Riccardo Betti, Soumyendu Sarkar. This is a U.S. Department of Energy (DoE)-funded collaborative work with the physicists at the University of Rochester. The research applied advanced sample-efficient AI techniques that leverage existing knowledge to drastically reduce the experiments required for optimizing energy yield in inertial confinement fusion, which has promise for the limitless clean energy.
“LLM Enhanced Bayesian Optimization for Scientific Applications like Fusion,” by Sahand Ghorbanpour, Ricardo Luna Gutierrez, Vineet Gundecha, Desik Rengarajan, Ashwin Ramesh Babu, Soumyendu Sarkar. This research integrates large language models (LLMs) with Bayesian Optimization (BO) to develop cost-effective, sample-efficient methods that outperform traditional approaches in optimizing complex costly scientific experiments such as Inertial Confinement Fusion (ICF). This is also funded by the U.S. DoE for AI research for fusion energy.
