Seminar

As capabilities of foundation models skyrocket, openness plummets. In this talk, I argue that open-source models are essential for the long-term goal of building a rigorous foundation for AI. Greater access---from API to open-weight to open-source---enables deeper forms of research.  API access allows us to push the frontier of agents, and I will present our recent work on simulation and problem-solving agents. Open weights enables reproducible research on safety, interpretability, and more generally, “model forensics”. Open-source unlocks fundamental innovations in architectures, training procedures, and data curation methods. Of course, the key obstacle for building open-source models is the resources required (data, compute, and research/engineering). I will conclude with some promising directions that leverage the community that bring us closer to the vision of open-source foundation models.

Bio: Percy Liang is an Associate Professor of Computer Science at Stanford University (B.S. from MIT, 2004; Ph.D. from UC Berkeley, 2011) and the director of the Center for Research on Foundation Models (CRFM). He is currently focused on making foundation models (in particular, language models) more accessible through open-source and understandable through rigorous benchmarking. In the past, he has worked on many topics centered on machine learning and natural language processing, including robustness, interpretability, human interaction, learning theory, grounding, semantics, and reasoning. He is also a strong proponent of reproducibility through the creation of CodaLab Worksheets. His awards include the Presidential Early Career Award for Scientists and Engineers (2019), IJCAI Computers and Thought Award (2016), an NSF CAREER Award (2016), a Sloan Research Fellowship (2015), a Microsoft Research Faculty Fellowship (2014), and paper awards at ACL, EMNLP, ICML, COLT, ISMIR, CHI, UIST, and RSS.Title: Open-Source and Science in the Era of Foundation ModelsAbstract: As capabilities of foundation models skyrocket, openness plummets. In this talk, I argue that open-source models are essential for the long-term goal of building a rigorous foundation for AI. Greater access---from API to open-weight to open-source---enables deeper forms of research.  API access allows us to push the frontier of agents, and I will present our recent work on simulation and problem-solving agents. Open weights enables reproducible research on safety, interpretability, and more generally, “model forensics”. Open-source unlocks fundamental innovations in architectures, training procedures, and data curation methods. Of course, the key obstacle for building open-source models is the resources required (data, compute, and research/engineering). I will conclude with some promising directions that leverage the community that bring us closer to the vision of open-source foundation models.

---

PLI YouTube

Please submit your questions in advance HERE.

Foundation models such as large language models (LLMs) and vision-language models (VLMs) have showcased remarkable capabilities in areas like commonsense reasoning, planning, and physical understanding. These advancements have paved the way for applying such models in robotics, enabling them to perceive their environment, generate and evaluate potential plans, and even command low-level actions. However, the integration of foundation models in robotics also raises significant challenges, particularly around ensuring the safety and alignment of robotic systems given the black-box nature of these models. This symposium brings together leading experts from both academia and industry to explore the opportunities and challenges of safely incorporating foundation models into robotic systems.

---

Registration is free but required.
REGISTER HERE

Open-source hardware holds a promise of sustaining the progress of the semiconductor industry in the age of heterogeneous computing. It can enable design reuse, foster design collaboration, and support workforce development. ESP is an open-source research platform for system-on-chip (SoC) design that combines a modular architecture and an agile design methodology. The ESP architecture simplifies the design and prototyping of heterogeneous chips with multiple RISC-V processor cores and dozens of loosely-coupled accelerators, all interconnected with a scalable network-on-chip. The ESP methodology promotes system-level design while accommodating different specification languages and design flows. ESP's capabilities have allowed a small team of mostly graduate students to realize two SoCs of growing complexity, each in the span of just a few months. Conceived as a heterogeneous system integration platform and developed through years of teaching at Columbia University, ESP is intrinsically suited to advance collaborative engineering across the open-source hardware community.

Bio: Luca Carloni is professor and chair of Computer Science at Columbia University in the City of New York.  He holds a Laurea Degree Summa cum Laude in Electronics Engineering from the University of Bologna, Italy, and the MS and PhD degrees in Electrical Engineering and Computer Sciences from the University of California, Berkeley.  His research interests include heterogeneous computing, system-on-chip platforms, embedded systems, and open-source hardware. He co-authored over one hundred and eighty refereed papers.  Luca received the NSF CAREER Award, the Alfred P. Sloan Research Fellowship, and the ONR Young Investigator Award. He is an IEEE Fellow.

---

Cosponsored by the Department of Electrical and Computer Engineering and the Department of Computer Science

To request accommodations for a disability please contact Donna Ghilino, dg3548@princeton.edu, at least one week prior to the event.

---

Contributions to and/or sponsorship of any event does not constitute departmental or institutional endorsement of the specific program, speakers or views presented.

The past 50 years has seen a dramatic increase in the amount of compute per person, in particular, those enabled by AI. Despite the positive societal benefits, AI technologies come with significant environmental implications. I will talk about the scaling trend and the carbon footprint of AI computing by examining the model development cycle, spanning data, algorithms, and system hardware. At the same time, we will consider the life cycle of system hardware from the perspective of hardware architectures and manufacturing technologies. I will highlight key efficiency optimization opportunities for cutting-edge AI technologies, from deep learning recommendation models to multi-modal generative AI tasks. To scale AI sustainably, we need to make AI and computing, more broadly, efficient and flexible. We must also go beyond efficiency and optimize across the life cycle of computing infrastructures, from hardware manufacturing to datacenter operation and end-of-life processing for the hardware. Based on the industry experience and lessons learned, my talk will conclude with important development and research directions to advance the field of computing in an environmentally-responsible and sustainable manner.

Bio: Carole-Jean Wu is a Director of Research Science at Meta. She is a founding member and a Vice President of MLCommons – a non-profit organization that aims to accelerate machine learning for the benefits of all. Dr. Wu also serves on the MLCommons Board as a Director, chaired the MLPerf Recommendation Benchmark Advisory Board, and co-chaired for MLPerf Inference. Prior to Meta/Facebook, She was a professor at ASU. She earned her M.A. and Ph.D. from Princeton and B.Sc. from Cornell.

Dr. Wu's expertise sits at the intersection of computer architecture and machine learning. Her work spans across datacenter infrastructures and edge systems, such as developing energy- and memory-efficient systems and microarchitectures, optimizing systems for machine learning execution at-scale, and designing learning-based approaches for system design and optimization. Dr. Wu's work has been recognized with several awards, including IEEE Micro Top Picks and ACM/IEEE Best Paper Awards. She was the Program Co-Chair of the Conference on Machine Learning and Systems (MLSys) in 2022, the Program Chair of the IEEE International Symposium on Workload Characterization (IISWC) in 2018, and the Editor for the IEEE MICRO Special Issue on Environmentally Sustainable Computing. She currently serves on the ACM SIGARCH/SIGMICRO CARES committee.

---

Cosponsored by the Department of Electrical and Computer Engineering and the Department of Computer Science

To request accommodations for a disability please contact Donna Ghilino, dg3548@princeton.edu, at least one week prior to the event.

Excited states of quantum systems are relevant for photochemistry, quantum dots, semiconductors and more, but remain extremely challenging to calculate by conventional computational methods. In recent years tools from machine learning have found useful application in computational quantum mechanics, especially in making ground state calculations much more accurate. In this talk I will present a new algorithm for estimating the lowest excited states of a quantum system which works extremely well with deep neural networks. The method has no free parameters and requires no explicit orthogonalization of the different states, instead transforming the problem of finding excited states of a given system into that of finding the ground state of an expanded system. Expected values of arbitrary observables can be calculated including off-diagonal expectations between different states such as the transition dipole moment. We show that by combining this method with the FermiNet and Psiformer Ansatze we can accurately recover vertical excitation energies and oscillator strengths on molecules as large as benzene. Beyond the examples on molecules presented here we expect this technique will be of great interest for applications to atomic, nuclear, and condensed matter physics.

Bio: David Pfau is a staff research scientist at Google DeepMind and a visiting professor at Imperial College London in the Department of Physics, where he supervises work on applications of deep learning to computational quantum mechanics. His research interests span artificial intelligence, machine learning and scientific computing.

Prior to joining DeepMind, he was a PhD student at the Center for Theoretical Neuroscience at Columbia, where I worked on algorithms for analyzing and understanding high-dimensional data from neural recordings with Liam Paninski and nonparametric Bayesian methods for predicting time series data with Frank Wood. He also had a stint as a research assistant in the group of Mike DeWeese at UC Berkeley, jointly between Physics and the Redwood Center for Theoretical Neuroscience. His current research interests include applications of machine learning to computational physics and connections between differential geometry and unsupervised learning.

PNI/CS Special Seminar

---

Human cognition is characterized by a remarkable ability to transcend the specifics of limited experience to entertain highly general, abstract ideas. Efforts to explain this capacity have long fueled debates between proponents of symbol systems and statistical approaches. In this talk, I will present an approach that suggests a novel reconciliation to this long-standing debate, by exploiting an inductive bias that I term the relational bottleneck. This approach imbues neural networks with key properties of traditional symbol systems, thereby enabling the data-efficient acquisition of cognitive abstractions, without the need for pre-specified symbolic representations. I will also discuss studies of perceptual decision confidence that illustrate the need to ground cognitive theories in the statistics of real-world data, and present evidence for the presence of emergent reasoning capabilities in large-scale deep neural networks (albeit requiring far more training data than is developmentally plausible). Finally, I will discuss the relationship of the relational bottleneck to other inductive biases, such as object-centric visual processing, and consider the potential mechanisms through which this approach may be implemented in the human brain.

Bio: Taylor Webb received his PhD in Cognitive Psychology and Neuroscience from Princeton University, where he studied with Michael Graziano and Jonathan Cohen. He is now a postdoctoral research fellow in the Psychology Department at UCLA, working with Hongjing Lu, Keith Holyoak, and Hakwan Lau. His research is focused on the question of how the brain extracts structured, abstract representations from noisy, high-dimensional perceptual inputs, and uses these representations to achieve intelligent behavior. To better understand these processes, his work exploits a bidirectional interaction between cognitive science and artificial intelligence. This involves both the use of techniques from AI to build models of higher-order cognitive processes (e.g., metacognition and analogical reasoning) that are grounded in realistic perceptual inputs (e.g., images and natural language), and the development of novel AI systems that take inspiration from cognitive science and neuroscience to achieve more human-like learning and reasoning.

---

To request accommodations for a disability please contact Yi Liu, irene.yi.liu@princeton.edu, at least one week prior to the event.

PNI/CS Special Seminar

---

The brain is an unparalleled learning machine, yet the principles that govern learning in the brain remain unclear. In this talk I will suggest that depth–the serial propagation of signals–may be a key principle sculpting learning dynamics in the brain and mind. To understand several consequences of depth, I will present mathematical analyses of the nonlinear dynamics of learning in a variety of simple solvable deep network models. Building from this theoretical work, I will trace implications for the development of human semantic cognition, showing that deep but not shallow networks exhibit hierarchical differentiation of concepts through rapid developmental transitions and ubiquitous semantic illusions between such transitions. Finally, turning to rodent systems neuroscience, I will show that deep network dynamics can account for individually variable yet systematic transitions in strategy as mice learn a visual detection task over several weeks. Together, these results provide analytic insight into how the statistics of an environment can interact with nonlinear deep learning dynamics to structure evolving neural representations and behavior over learning.

Bio: Andrew Saxe is a Professorial Research Fellow at the Gatsby Computational Neuroscience Unit and Sainsbury Wellcome Centre at UCL. He was previously an Associate Professor in the Department of Experimental Psychology at the University of Oxford. His research focuses on the theory of deep learning and its applications to phenomena in neuroscience and psychology. His work has been recognized by the Robert J. Glushko Dissertation Prize from the Cognitive Science Society, the Blavatnik UK Finalist Award in Life Sciences, and a Schmidt Science Polymath award. He is also a CIFAR Azrieli Global Scholar in the CIFAR Learning in Machines & Brains programme.

---

To request accommodations for a disability please contact Yi Liu, irene.yi.liu@princeton.edu, at least one week prior to the event.

Computing has a substantial environmental impact, with far-reaching consequences. Information and Communication Technologies (ICT) alone contribute to 4% of global carbon emissions, a figure equivalent to the aviation industry's carbon footprint. This impact is poised to increase as the demand for computing continues to surge, particularly at the edge and within massive hyperscale data centers. Creating environmentally sustainable computing solutions presents unique challenges that go beyond energy efficiency optimization. It necessitates a holistic approach that considers various factors. For example, sustainable computing devices must account for the carbon emissions generated during the manufacturing process, the influence of renewable energy sources that fluctuate by location and time on a system's operational emissions, and the intricate interplay between embodied and operational emissions. In this presentation, we will delve into recent breakthroughs and the research challenges we confront as we strive to transition to environmentally responsible computing systems. With a primary focus on the carbon footprint of computing, we will explore recent initiatives aimed at characterizing and comprehending the environmental impact of computing. We will also introduce new carbon accounting models and frameworks, highlighting their application in pertinent domains, such as AI, to promote the development of sustainable AI solutions.

Bio: Udit Gupta is an Assistant Professor at Cornell Tech, where his research is dedicated to pioneering the development of cutting-edge and ethically responsible AI platforms through the innovation of novel computer systems and hardware. His recent endeavors are centered around the design and optimization of data center-based, deep learning-driven personalized recommendation systems and championing eco-friendly computing by meticulously considering the ecological implications associated with the entire hardware life cycle.

Udit's groundbreaking research has been tested and validated within production data centers at a large scale but has also found its place in established benchmarks and infrastructures that are widely embraced by the research community. His contributions were acknowledged with an honorable mention as an IEEE MICRO Top Picks selection in 2020 and further distinguished with an IEEE MICRO Top Picks award in 2021 and 2022. His work earned nominations for the Best Paper awards at PACT 2019 and DAC 2018. Furthermore, his doctoral research received the prestigious SIGARCH Outstanding Ph.D. Dissertation award honorable mention and the SIGMICRO Outstanding Ph.D. Dissertation award honorable mention.

As the impact of artificial intelligence (AI) continues to proliferate, computer architects must assess and mitigate its environmental impact. This talk will survey strategies for reducing the carbon footprint of AI computation and datacenter infrastructure, drawing on data and experiences from industrial, hyperscale systems. First, we analyze the embodied and operational carbon implications of super-linear AI growth. Second, we re-think datacenter infrastructure and define a solution space for carbon-free computation with renewable energy, utility-scale batteries, and job scheduling. Finally, we develop strategies for datacenter demand response, incentivizing both batch and real-time workloads to modulate power usage in ways that reflect their performance costs. In summary, the talk provides a broad perspective on sustainable computing and outlines the many remaining directions for future work.

Bio: Benjamin C. Lee is a Professor of Electrical and Systems Engineering and of Computer and Information Science at the University of Pennsylvania. Dr. Lee’s research focuses on computer architecture, energy efficiency, environmental sustainability, and security. He builds interdisciplinary links to machine learning and algorithmic game theory to better design and manage computer systems. His research has been recognized by IEEE Micro Top Picks, Communications of the ACM Research Highlights, as well as publication honors from the ASPLOS, HPCA, MICRO, and SC conferences. Dr. Lee was an Assistant and then Associate Professor at Duke University. He received his post-doctorate in Electrical Engineering at Stanford University, Ph.D. in Computer Science from Harvard University, and B.S. in Electrical Engineering and Computer Science from the University of California at Berkeley. He has also held visiting positions at Meta AI, Microsoft Research, Intel Labs, and Lawrence Livermore National Lab. Dr. Lee received the NSF Computing Innovation Fellowship, NSF CAREER Award, and Google Faculty Research Award. He is an ACM Distinguished Scientist and IEEE Senior Member.

---

This seminar is hosted by the department of Electrical and Computer Engineering.

Please REGISTER HERE
Reception following talk.

---

While public blockchains like Bitcoin and Ethereum are celebrated for their security, and decentralized nature, they're built on the insecure and best-effort  structures of the Internet. In this talk, we dive into the often overlooked issues arising from the Internet layer, uncovering how centralization and certain assumptions can make blockchains vulnerable. Bad actors can exploit these weak spots, to hinder blockchain functionality, compromise security and user privacy. Beyond these challenges, we'll also spotlight the growing centralization concerns and their implications. But it's not all cautionary tales! We'll share deployable solutions and strategies to fortify blockchains against these threats.

Bio: Maria Apostolaki, assistant professor of electrical and computer engineering, joined the Princeton faculty in August 2022 after a one-year postdoctoral position at Carnegie Mellon University. She earned her Ph.D. from ETH Zurich in 2021. Her research draws from networking, security, blockchain, and machine learning. Overall, her goal is to design and build networked systems that are secure, reliable and performant. She was named a Rising Star in Computer Networking and Communications by N2Women. She is a recipient of the Google Research Scholar Award recognizing her work on improving network monitoring.