- talks - lists.cs.princeton.edu

Jane Pan will present her MSE talk "What In-Context Learning “Learns” In-Context: Disentangling Task Recognition and Task Learning" on Monday, April 24, 2023 at 2pm in CS 402.
by Nicki Mahler 21 Apr '23

21 Apr '23

Jane Pan will present her MSE talk "What In-Context Learning “Learns” In-Context: Disentangling Task Recognition and Task Learning" on Monday, April 24, 2023 at 2pm in CS 402. The members of her committee are as follows: Advisor: Danqi Chen; Reader: Karthik Narasimhan Title: What In-Context Learning “Learns” In-Context: Disentangling Task Recognition and Task Learning Abstract: Large language models (LLMs) exploit in-context learning (ICL) to solve tasks with only a few demonstrations, but its mechanisms are not yet well-understood. Some works suggest that LLMs only recall already learned concepts from pre-training, while others hint that ICL performs implicit learning over demonstrations. We characterize two ways through which ICL leverages demonstrations: task recognition (TR) captures the extent to which LLMs can recognize a task through demonstrations – even without ground-truth labels – and apply their pre-trained priors; task learning (TL) is the ability to capture new input- label mappings unseen in pre-training. Using a wide range of classification datasets and two LLM families (GPT-3 and OPT), we design controlled experiments to disentangle the roles of TR and TL in ICL. We show that (1) models can achieve non-trivial performance with only TR, and TR does not further improve with larger models or more demonstrations; (2) LLMs acquire TL as the model size scales; and (3) the correlation between model sizes and TL performance strengthens with the number of demonstrations. Our findings unravel two dif- ferent forces behind ICL and we advocate for discriminating them in future ICL research due to their distinct nature.

1 1

CS Department Colloquium Series: week of Mar 27-31
by Emily C. Lawrence 21 Apr '23

21 Apr '23

Here are next week's CS Department Colloquium Series talks. As always, you can find the full calendar of events here: https://www.cs.princeton.edu/general/newsevents/events Speaker: Adam Fisch, Massachusetts Institute of Technology Date: Monday, March 27 Time: 12:30pm EST Location: CS 105 Host: Karthik Narasimhan Event page: https://www.cs.princeton.edu/events/26363 Title: Towards Efficient and Reliable Machine Learning for Natural Language Processing (and Beyond) Abstract: In this talk, I will introduce work on fundamental techniques for building and deploying effective natural language processing (NLP) systems that are also efficient and reliable. Specifically, I will address three interconnected challenges for modern machine learning in NLP: how to quickly adapt foundation models to new tasks with limited data, how to dynamically reconfigure large architectures for more efficient computation, and how to develop powerful theoretical tools for rigorous, yet practical, uncertainty quantification. To conclude, I will highlight a number of my future research directions, as well as extensions to interesting applications beyond natural language. Bio: Adam Fisch is a PhD candidate at MIT working with Regina Barzilay and Tommi Jaakkola, and a recipient of an NSF Graduate Research Fellowship. His research centers around principled methods for efficient and reliable machine learning systems that work effectively in realistic scenarios, and has appeared in top-tier venues such as *ACL, ICLR, ICML, and NeurIPS. Adam also served as a co-instructor for the tutorial on Uncertainty Estimation for NLP at COLING 2022, and as a co-organizer of the Machine Reading for Question Answering workshops at EMNLP 2019 and 2021. Prior to MIT, Adam was a research engineer at Meta (Facebook) AI Research for two years, and studied mechanical engineering as an undergraduate at Princeton University. Speaker: Aviral Kumar, University of California, Berkeley Date: Tuesday, March 28 Time: 12:30pm EST Location: CS 105 Host: Jia Deng Event page: https://www.cs.princeton.edu/events/26377 Title: Reinforcement Learning from Static Datasets: Algorithms, Analysis and Applications Abstract: Typically, reinforcement learning (RL) methods rely on trial-and-error interaction with the environment from scratch to discover effective behaviors. While this sort of paradigm has the potential to discover good strategies, this paradigm also inhibits RL methods from collecting enough experience or training data in real-world problems where active interaction is expensive (e.g., in drug design) or dangerous (e.g., for robots operating around humans). My work develops approaches to alleviate this limitation: how can we learn policies to effectively make decisions entirely from previously-collected, static datasets in an offline manner? In this talk, I will discuss challenges that appear in this kind of offline reinforcement learning (offline RL) and develop algorithms and techniques to address these challenges. I will then discuss how my approaches for offline RL and decision-making have enabled us to make progress in real-world problems such as hardware accelerator design, robotic manipulation, and computational chemistry. Finally, I will discuss how we can enable offline RL methods to benefit from generalization capabilities offered by large and expressive models, similar to supervised learning. Bio: Aviral Kumar is a final year Ph.D. student at UC Berkeley. His research focuses on developing effective and reliable approaches for (sequential) decision-making. Towards this goal, he focuses on designing reinforcement learning techniques to static datasets and on understanding and applying these methods in practice. Before his Ph.D., Aviral obtained his B.Tech. in Computer Science from IIT Bombay in India. He is a recipient of the C.V. & Daulat Ramamoorthy Distinguished Research Award, given to 1 PhD student in EECS at Berkeley for outstanding contributions to a new area of research in computer science, Facebook Ph.D. Fellowship in Machine Learning and Apple Scholars in AI/ML Ph.D. Fellowship.

1 5

Alfredo Velasco will present his talk "Identifying Cancer-Driver Mutations using Machine Learning" on Tuesday, April 25, 2023 at 2pm in FCHEM 181
by Gradinfo 20 Apr '23

20 Apr '23

Alfredo Velasco will present his MSE talk "Identifying Cancer-Driver Mutations using Machine Learning" on Tuesday, April 25, 2023 at 2pm in FCHEM 181. The members of his committee are as follows: Mona Singh (adviser) and Yuri Pritykin (reader) All are welcome to attend. Talk title and abstract follow below. Title: "Identifying Cancer-Driver Mutations using Machine Learning" Abstract: Cancer is a cellular disease which is caused by somatic alterations that result in the increased growth rate of a cell. Many of these alterations are mutations that occur within protein-coding regions of the genome. Cancerous cells will also contain mutations that don't contribute to the disease progression. Thus, there is a need to develop methods that can differentiate between mutations that cause cancer (called driver mutations) and mutations that don't (called passenger mutations). This project specialized in looking at different methods to generate a set of features that could be fed into a Machine Learning (ML) model. One of these methods of feature generation includes protein language models which result in high-dimensional representations of amino acids that capture the context within which they appear within protein sequences. Another more familiar method for feature generation is to obtain the physiochemical properties of the amino acids such as the hyrdrophobic properties or if they are evolutionarily conserved. We also look at a variety of ML models such as random forests, gradient boosting, Logistic Regression, and Gaussian Naive Bayes, and Decision Tree to determine which model gives the best prediction. Together, the ultimate goal of this project is to determine which combination of feature sets and ML models give the best prediction performance and can properly distinguish driver and passenger mutations.

1 0

John Yang will present his MSE talk "Towards an Enhanced, Faithful, and Adaptable Web Interaction Environment" on Monday, April 24, 2023 at 3pm in CS 301.
by Gradinfo 20 Apr '23

20 Apr '23

John Yang will present his MSE talk "Towards an Enhanced, Faithful, and Adaptable Web Interaction Environment" on Monday, April 24, 2023 at 3pm in CS 301. The members of his committee are as follows: Karthik Narasimhan (adviser) and Danqi Chen (reader) All are welcome to attend. Talk title and abstract follow below. Title: Towards an Enhanced, Faithful, and Adaptable Web Interaction Environment Abstract: We identify key areas of improvement for WebShop, an e-commerce shopping environment for training decision making language agents. Specifically, shortcomings in: 1) faithfulness of the reward function to human evaluation, 2) comprehensiveness of its content, and 3) human participation required for generating instructions has hindered WebShop’s promises to be a scalable real-world environment. To solve these issues, we first incorporate greater faithfulness to human evaluation by designing a new reward function to capture lexical similarities and synonyms. Second, we identify customer reviews, similar products, and customer FAQs as missing semantic components that are most helpful to human execution of the task from surveying 75 respondents. Finally, we reformulate the attribute tagging problem as a extractive short-phrase prediction task to enhance scalability. Our V2 reward function closes the gap between the scores of the WebShop’s automated reward function (from 81.5% to 87.7%) and human evaluation (89.9%). Our attribute tagging approach achieves an accuracy of 72.2% with a t5-3b model fine tuned on 2, 000 training data points, showing potential to automate the instruction creation pipeline.

1 0

Kathryn Wantlin will present her MSE talk "Spatial Birth-Death Point Process Models for Controllable Morphogenesis" on Thursday, April 20, 2023 at 2pm in CS 301.
by Gradinfo 20 Apr '23

20 Apr '23

Kathryn Wantlin will present her MSE talk "Spatial Birth-Death Point Process Models for Controllable Morphogenesis" on Thursday, April 20, 2023 at 2pm in CS 301. The members of her committee are as follows: Ryan Adams (adviser), and Naomi Leonard (MAE, reader) All are welcome to attend. Title and abstract follow below. Title: Spatial Birth-Death Point Process Models for Controllable Morphogenesis Abstract: Controlling morphogenetic processes - where the total mass of agents varies as the system evolves - is a particularly difficult inverse design problem. In this thesis, we present a model for such bio-inspired shape formation using spatial point processes, a popular framework for describing many types of spatio-temporal reproductive processes. We frame the problem as one of controlling spatial birth-death processes, and given a target stationary distribution for the cell-like points, construct a continuous-time Markov chain Monte Carlo (MCMC) procedure by which the points reproduce to achieve and maintain the target shape. Within this framework, we consider several types of point processes, beginning with the spatially uncorrelated Poisson process and extending to the repulsive Strauss process, demonstrating how the combination of global and local information allows the points to determine their relative birth-death rates. To improve biological plausibility, we also extend the procedure to include reversible split-and-merge transitions and spatial Hawkes-based birth-death rates, thereby modeling the locality of cell division. This stochastic, continuous-time MCMC formulation of the problem possesses desirable properties for multi-agent systems, including recovery from partial destruction via ergodicity of the constructed Markov chain, and enables extensions into marked spatial point processes and periodic patterns.

1 0

Princeton Robotics Seminar: Renee Zhao, Friday April 21
by Emily C. Lawrence 19 Apr '23

19 Apr '23

Speaker: Renee Zhao, Assistant Professor of Mechanical Engineering, Stanford University Date: Friday, April 21, 2023 Time: 11am eastern time Location: Computer Science 105 Event page: [ https://robo.princeton.edu/seminar | https://robo.princeton.edu/seminar ] Title: Multifunctional Origami Robots Abstract: In this talk, I will introduce our recent work on origami mechanisms and actuation strategies for applications spanning from biomedical devices to foldable space structures. The first topic is magnetically actuated millimeter-scale origami medical robots for effective amphibious locomotion in severely confined spaces or aqueous environments. The origami robots are based on the Kresling origami, whose thin shell structure 1) provides an internal cavity for drug storage, 2) permits torsion-induced contraction as a crawling mechanism and a pumping mechanism for controllable liquid medicine dispensing, 3) serves as propellers that spin for propulsion to swim, 4) offers anisotropic stiffness to overcome the large resistance from the severely confined spaces in biomedical environments. For the second part of my talk, the concept of hexagonal ring origami folding mechanism will be introduced as a strategy for deployable/foldable structures for space applications. The hexagonal rings can tessellate 2D/3D surfaces and each ring can snap to its stable folded configuration with only 10.6% of the initial area. Through finite-element analysis and the rod model, snap-folding of the hexagonal ring with slight geometric modification and residual strain are studied for easy folding of the ring to facilitate the design and actuation of hexagonal ring origami assemblies for functional foldable structures with extreme packing ratio. Bio: Renee Zhao is an Assistant Professor of Mechanical Engineering at Stanford University. Renee received her PhD degree in Solid Mechanics from Brown University in 2016. She spent two years as a postdoc associate at MIT working on modeling of soft composites. Before Renee joined Stanford, she was an Assistant Professor at The Ohio State University from 2018 to 2021. Renee’s research concerns the development of stimuli-responsive soft composites and shape morning mechanisms for multifunctional robotic systems. Renee is a recipient of the NSF Career Award (2020), AFOSR YIP (2023), ASME Journal of Applied Mechanics award (2021), the 2022 ASME Pi Tau Sigma Gold Medal, and the 2022 ASME Henry Hess Early Career Publication Award.

1 0

Ryan Torok will present his MSE talk "Leveraging Sandboxes for a Minimally Invasive Browser Fingerprinting Defense" on Tuesday, April 25, 2023 at 2pm in CS 401.
by Nicki Mahler 18 Apr '23

18 Apr '23

Ryan Torok will present his MSE talk "Leveraging Sandboxes for a Minimally Invasive Browser Fingerprinting Defense" on Tuesday, April 25, 2023 at 2pm in CS 401. Advisor: Amit Levy Reader: Jonathan Mayer Title: Leveraging Sandboxes for a Minimally Invasive Browser Fingerprinting Defense Abstract: We present Sandcastle, an entropy-based browser fingerprinting defense that aims to minimize its interference with legitimate web applications. Sandcastle allows developers to partition code that operates on identifiable information into sandboxes to prove to the browser the information cannot be sent in any network request. Meanwhile, sandboxes may make full use of identifiable information on the client side, including writing to dedicated regions of the Document Object Model. For applications where this policy is too strict, Sandcastle provides an expressive cashier that allows precise control over the granularity of data that is leaked to the network. These features allow Sandcastle to eliminate most or all of the noise added to the outputs of identifiable APIs by Chrome’s Privacy Budget framework, the current state of the art in entropy-based fingerprinting defenses. Enabling unlimited client-side use of identifiable information allows for a much more comprehensive set of web applications to run under a fingerprinting defense, such as 3D games and video streaming, and provides a mechanism to expand the space of APIs that can be introduced to the web ecosystem without sacrificing privacy.

1 0

CITP Lecture: Peter Henderson, Tuesday, April 18 at 4:30pm
by Emily C. Lawrence 14 Apr '23

14 Apr '23

CITP Lecture Speaker: Peter Henderson, Stanford University Date: Tuesday, April 18 Time: 4:30pm EST Location: CS 105 Host: Center for Information Technology Policy (CITP) Event page: https://citp.princeton.edu/event/citp-lecture-henderson/ Title: Aligning Machine Learning, Law, and Policy for Responsible Real-World Deployments Abstract: Machine learning (ML) is being deployed to a vast array of real-world applications with profound impacts on society. ML can have positive impacts, such as aiding in the discovery of new cures for diseases and improving government transparency and efficiency. But it can also be harmful: reinforcing authoritarian regimes, scaling the spread of disinformation, and exacerbating societal biases. As we rapidly move toward systemic use of ML in the real world, there are many unanswered questions about how to successfully use ML for social good while preventing its potential harms. Many of these questions inevitably require pursuing a deeper alignment between ML, law, and policy. Are certain algorithms truly compliant with current laws and regulations? Is there a better design that can make them more tuned to the regulatory and policy requirements of the real world? Are laws, policies, and regulations sufficiently informed by the technical details of ML algorithms, or will they be ineffective and out-of-sync? In this talk, we will discuss ways to bring together ML, law, and policy to address these questions. We will draw on real-world examples throughout the talk, including a unique real-world collaboration with the Internal Revenue Service. We will show how investigating questions of alignment between ML, law, and policy can advance core research in ML, drive interdisciplinary work, and reimagine how we think about certain laws and policies. It is Henderson’s hope that this agenda will help us lead to more effective and responsible ways of deploying ML in the real world, so that we steer toward positive impacts and away from potential harms. Bio: Peter Henderson is a joint J.D.-Ph.D. (Computer Science, AI) candidate at Stanford University where he is advised by Dan Jurafsky for his Ph.D. and Dan Ho for my J.D. at Stanford Law School. He is also an OpenPhilanthropy AI Fellow and a Graduate Student Fellow at the Regulation, Evaluation, and Governance Lab. At Stanford Law School. Henderson co-led the Domestic Violence Pro Bono Project, worked on client representation with the Three Strikes Project, and contributed to the Stanford Native Law Pro Bono Project. Previously, he was advised by David Meger and Joelle Pineau for his M.Sc. at McGill University and the Montréal Institute for Learning Algorithms. Henderson has spent time as a software engineer and applied scientist at Amazon AWS/Alexa and worked with Justice Cuéllar at the California Supreme Court. He is a part-time researcher with the Internal Revenue Service Research, Applied Analytics and Statistics Division, a technical advisor at the Institute for Security+Technology. His research focuses on aligning machine learning, law, and policy for responsible real-world deployments. This alignment process is two-fold: (1) guided by law, policy and ethics, develop general AI systems capable of safely tackling longstanding challenges in government and society; (2) empowered by a deep technical understanding of AI, make sure that laws & policies keep general AI systems safe and beneficial for all. Some of his work has received coverage by TechCrunch, Science, The Wall Street Journal, Bloomberg, and more. He also occasionally posts a roundup of news at the intersection of AI, Law, and Policy, the latest of which is here. Overall, he is interested in a wide range of other technical machine learning research, policy, and legal work.

1 0

Yushan Su will present her FPO "Making Neural Network Models More Efficient" on Wednesday, May 10, 2023 at 10am in Friend 202.
by Nicki Mahler 13 Apr '23

13 Apr '23

Yushan Su will present her FPO " Making Neural Network Models More Efficient" on Wednesday, May 10, 2023 at 10am in Friend 202. The members of her committee are as follows: Examiners: Kai Li (adviser), Wyatt Lloyd, and Karthik Narasimhan; Readers: Ravi Netravali and Olga Troyanskaya. Please see title and abstract below. All are welcome to attend. Complex machine learning tasks typically require large neural network models. However, training and inference on neural models require substantial compute power and large memory footprints, and incur significant cost. My thesis studies methods to make neural networks efficient at a relatively low cost. First, we explore how to utilize CPU servers for training and inference. CPU servers are more readily available, have larger memories, and cost less than GPUs or hardware accelerators. However, they are much less efficient for training and inference tasks. My thesis studies how to design efficient software kernels for sparse neural networks that allow unstructured pruning to achieve efficiency of training or inference with minimum accuracy losses. Our evaluation shows that our sparse kernels can achieve 6.4x-20.0x speedups at different sparsities over the commonly used Intel MKL sparse library. Second, we study how to achieve high-throughput inference for large models. We explore ways to combine data multiplexing with pruning. We find that combining the two approaches can achieve better throughput than using each approach alone for a given accuracy loss. We then propose an automatic method to find parameters for such combinations to maximize inference throughput given an accuracy loss budget. We show that the proposed combining can improve Transformer models throughput by 7.5-29.5X at different accuracy thresholds, which is higher than using either approach alone.

1 2

From Big Data To Equitable Health: Princeton Precision Health Spring Workshop May 2
by Emily C. Lawrence 13 Apr '23

13 Apr '23

Please save May 2, 2023, from 9:30 am - 5:00 pm, for the first Princeton Precision Health (PPH) workshop on interdisciplinary, date-driven approaches for precision medicine. We will bring together biomedical researchers, clinicians, computer scientists, ethicists, and social scientists to discuss a wide range of topics, working toward our goal of enabling more responsive, equitable, and precisely targeted health care. Registration is free but required for an accurate head count for breakfast and lunch: [ https://forms.gle/XFa4RSNhcrDTo5gp7 | https://forms.gle/XFa4RSNhcrDTo5gp7 ] Confirmed participants include: Matthias Kretzler, MD, Warner-Lambert/Parke-Davis Professor, Internal Medicine, University of Michigan. Sridhar Venkatapuram, Senior Lecturer (Associate Professor) in Global Health and Philosophy. Deputy Director, Global Health Institute. Director of Global Health Education and Training. King’s College London. Mona Singh, Wang Family Professor in Computer Science. Professor of Computer Science and the Lewis-Sigler Institute for Integrative Genomics, Princeton University. Coleen Murphy, James A. Elkins Professor in the Life Sciences. Professor of Molecular Biology and the Lewis-Sigler Institute for Integrative Genomics, Director of the Paul Glenn Laboratories for Aging Research at Princeton University. Ian Peebles, Harold T. Shapiro Postdoctoral Research Associate in Bioethics at the Center for Human Values at Princeton University. David Kinney, Postdoctoral Research Associate in Cognitive Science of Values at the Center for Human Values at Princeton University. Emily Foster-Hanson, Postdoctoral Research Associate in Cognitive Science of Values at Princeton University (Current); Assistant Professor of Psychology at Swarthmore College (Fall 2023). Hope to see you there! Olga Troyanskaya *************************************************************** Olga Troyanskaya Director, Princeton Precision Health Professor of Computer Science and the Lewis-Sigler Institute for Integrative Genomics Princeton University

1 1