April 2023 - talks - lists.cs.princeton.edu

Vikram Ramaswamy will present his FPO "Tackling bias within Computer Vision Models" on Monday, May 8, 2023 at 12:00 PM in CS 402.
by Louis W. Riehl 26 Apr '23

26 Apr '23

Vikram Ramaswamy will present his FPO "Tackling bias within Computer Vision Models" on Monday, May 8, 2023 at 12:00 PM in CS 402. Location: CS 402 The members of Vikram’s committee are as follows: Examiners: Olga Russakovsky (Adviser), Ellen Zhong, Ryan Adams Readers: Jia Deng, Andrés Monroy-Hernández A copy of his thesis is available upon request. Please email gradinfo(a)cs.princeton.edu if you would like a copy of the thesis. Everyone is invited to attend his talk. Abstract follows below: Over the past decade the rapid increase in the ability of computer vision models has led to their applications in a variety of real-world applications from self-driving cars to medical diagnoses. However, there is increasing concern about the fairness and transparency of these models. In this thesis, we tackle these issue of bias within these models along two different axes. First, we consider the datasets that these models are trained on. We use two different methods to create a more balanced training dataset. First, we create a synthetic balanced dataset by sampling strategically from the latent space of a generative network. Next, we explore the potential of creating a dataset through a method other than scraping the internet: we solicit images from workers around the world, creating a dataset that is balanced across different geographical regions. Both techniques are shown to help create models with less bias. Second, we consider methods to improve interpretability of these models, which can then reveal potential biases within the model. We investigate a class of interpretability methods called concept-based methods that output explanations for models in terms of human understandable semantic concepts. We demonstrate the need for more careful development of the datasets used to learn the explanation as well as the concepts used within these explanations. We construct a new method that allows for users to select a trade-off between the understandability and faithfulness of the explanation. Finally, we discuss how methods that completely explain a model can be developed, and provide heuristics for the same. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Gianluca Bencomo will present his MSE talk "Bayesian Filtering for Neural Networks" on Wednesday, April 26, 2023 at 9am in CS 401.
by Gradinfo 24 Apr '23

24 Apr '23

Gianluca Bencomo will present his MSE talk "Bayesian Filtering for Neural Networks" on Wednesday, April 26, 2023 at 9am in CS 401. The members of his committee are as follows: Tom Griffiths (adviser) and Ryan Adams (reader). All are welcome to attend. Please see talk details below. Title: Bayesian Filtering for Neural Networks Abstract: Adaptability is a crucial component of intelligent systems. In dynamic environments, biological organisms continuously adjust their behavior to match the demands of the current situation. However, this fundamental feature of biological intelligence conflicts with a typical assumption of machine learning, which assumes that the data generating process is static over time. Bayesian filtering offers a rich toolkit for breaking this assumption, but the complex and high-dimensional space of neural networks poses several challenges. This thesis investigates these problems and presents progress for handling them. We demonstrate that human-like paradigms, such as continual learning and meta-learning, can be naturally interpreted within the Bayesian filtering framework. Furthermore, we show that sequential empirical risk minimization (ERM) is an implicit Bayesian filter, which explains its superior performance in time-varying prediction problems compared to more sophisticated methods. By making this implicit filter explicit, we can design machine learning systems that can more effectively adapt and generalize across time. Although this work is still in progress, this thesis provides a promising start to addressing this problem and sets the foundation for future research in this important domain.

1 0

TOMORROW 4/25/2023 - ECE Seminar Guillermo Sapiro, Duke University and Apple
by Emily C. Lawrence 24 Apr '23

24 Apr '23

ECE SEMINAR Speaker: Guillermo Sapiro, Duke University and Apple Title: Health and Wellbeing Inspiring Foundational Machine learning and Image Analysis Day: Tuesday, April 25, 2023 Time: 4:30 PM Location: B205 Engineering Quadrangle Host: James Sturm Abstract: Health and wellbeing bring new challenges to ML, computer vision, and image analysis. We first describe how health inspired us to new definitions of fairness and robustness in ML, which we formally analyzed in health-relevant scenarios ranging from fairness without demographics to fairness in federated settings. Then, we explain how to extend classical econometric tools, in particular regression discontinuity, to study the causal effects of a behavioral health nudge in one of the largest observational studies of its kind. We close the core of the presentation by showing how to use physiological knowledge in tandem with ML to predict environmental-affected vital signs via a hybrid personalized modeling approach. To conclude, we showcase novel foundational models and techniques with a brief sample of other recent activities in the team, ranging from FDA-cleared deep brain stimulation target localization to behavioral/brain understanding in developmental and neurodegenerative disorders. The work presented is the result of close collaboration between clinicians, domain experts, engineers, computer scientists, app developers, mathematicians, and regulatory consultants. Bio: Guillermo Sapiro received his B.Sc. (summa cum laude), M.Sc., and Ph.D. from the Technion, Israel Institute of Technology. After post-doctoral research at MIT, he became Member of Technical Staff at HP Labs. He was with the Department of Electrical and Computer Engineering at the University of Minnesota. Currently he is a James B. Duke School Professor with Duke University. He is also a Distinguished Engineer with Apple, where he leads the Health AI team. He was awarded the ONR Young Investigator Award in 1998, the Presidential Early Career Awards for Scientist and Engineers (PECASE) in 1998, the NSF Career Award in 1999, and the National Security Science and Engineering Faculty Fellowship in 2010. He received the Test-of-Time awards at ICCV 2011 and at ICML 2019. He was elected to the American Academy of Arts and Sciences in 2018 and to the National Academy of Engineering in 2022, and is a Fellow of both IEEE and SIAM. G. Sapiro was the founding Editor-in-Chief of the SIAM Journal on Imaging Sciences.

1 0

Yiming Zuo will present his General Exam "Efficient View Synthesis with Point-based Representation" on Wednesday, April 26, 2023 at 3:00 PM in CS 302.
by Louis W. Riehl 21 Apr '23

21 Apr '23

Yiming Zuo will present his General Exam "Efficient View Synthesis with Point-based Representation" on Wednesday, April 26, 2023 at 3:00 PM in CS 302. Committee Members: Jia Deng (advisor), Szymon Rusinkiewicz, Felix Heide Abstract: We address the task of view synthesis, generating novel views of a scene given a set of images as input. In many recent works such as NeRF, the scene geometry is parameterized using neural implicit representations (i.e., MLPs). Implicit neural representations have achieved impressive visual quality but have drawbacks in computational efficiency. NeRF needs to evaluate the neural network at hundreds of 3D points along the ray, which is wasteful because most of the 3D spaces are unoccupied. NeRF’s implicit representation also makes it inflexible for scene editing operations such as deformation, which is important for downstream applications including augmented reality and video games. An intriguing question is whether we can achieve state-of-the-art visual quality by using explicit representations such as point clouds. The basic framework of point-based neural rendering is to represent the scene as a featurized point cloud. Although this framework has been studied in several recent works (Aliev et al., 2020; Wiles et al., 2020; Lassner & Zollhofer, 2021), the overall rendering quality still lags behind NeRF. Our approach adopts this basic framework but introduces a novel technique we call “Sculpted Neural Points (SNP)”, which significantly improves the robustness to the errors and holes in the reconstructed point cloud. We further propose a few novel designs in the point-based rendering pipeline that boost the performance. Compared to previous works that use point cloud-based representation, ours is the first model that achieves better rendering quality than NeRF, while being 100x faster in rendering, and reducing the training time by 66%. Finally, we show that our model allows fine-grained scene editing in a user-friendly way. Reading List: [ https://docs.google.com/document/d/15U9PNheu3dgfYe-_YZVA5vkLfVsYAwzGz2GDMDl… | https://docs.google.com/document/d/15U9PNheu3dgfYe-_YZVA5vkLfVsYAwzGz2GDMDl… ] Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so. Louis Riehl Graduate Administrator Computer Science Department, CS213 Princeton University (609) 258-8014

1 0

Re: [talks] CORRECTION Xinhao Liu will present his General Exam "Alignment Problems in Spatially Resolved Transcriptomics Data" on Wednesday, April 26, 2023 at 2:00 PM in CS 401.
by Louis Riehl 21 Apr '23

21 Apr '23

The time and location in the subject line (but not the main body) of the previous announcement were incorrect. Both are corrected here. Sorry for the confusion. From: talks-bounces(a)lists.cs.princeton.edu <talks-bounces(a)lists.cs.princeton.edu> On Behalf Of Louis W. Riehl Sent: Friday, April 21, 2023 2:00 PM To: talks <talks(a)lists.cs.princeton.edu> Subject: [talks] Xinhao Liu will present his General Exam "Alignment Problems in Spatially Resolved Transcriptomics Data" on Wednesday, April 26, 2023 at 3:00 PM in CS 302. Xinhao Liu will present his General Exam "Alignment Problems in Spatially Resolved Transcriptomics Data" on Wednesday, April 26, 2023 at 2:00 PM in CS 401. Committee Members: Ben Raphael (advisor), Yuri Pritykin, Olga Troyanskaya Abstract: Spatially resolved transcriptomics (SRT) technologies measure mRNA expression at thousands of locations in a tissue slice. However, nearly all SRT technologies measure expression in two dimensional slices extracted from a three-dimensional tissue, thus losing information that is shared across multiple slices from the same tissue. Integrating SRT data across multiple slices can help recover this information and improve downstream expression analyses, but multi-slice alignment and integration remains a challenging task. Existing methods for integrating SRT data either do not use spatial information or assume that the morphology of the tissue is largely preserved across slices, an assumption that is often violated due to biological or technical reasons. We introduce PASTE2, a method for partial alignment and 3D reconstruction of multi-slice SRT datasets, allowing only partial overlap between aligned slices and/or slice-specific cell types. PASTE2 formulates a novel partial Fused Gromov-Wasserstein Optimal Transport problem, which we solve using a conditional gradient algorithm. PASTE2 includes a model selection procedure to estimate the fraction of overlap between slices, and optionally uses information from histological images that accompany some SRT experiments. We show on both simulated and real data that PASTE2 obtains more accurate alignments than existing methods. We further use PASTE2 to reconstruct a 3D map of gene expression in a Drosophila embryo from a 16 slice Stereo-seq dataset. PASTE2 produces accurate alignments of multi-slice datasets from multiple SRT technologies, enabling detailed studies of spatial gene expression across a wide range of biological applications Reading List: https://docs.google.com/document/d/1JhXxEUxx8JpUv_wLDHTp44LG-cg0EKcLVX647y5… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

Xinhao Liu will present his General Exam "Alignment Problems in Spatially Resolved Transcriptomics Data" on Wednesday, April 26, 2023 at 3:00 PM in CS 302.
by Louis W. Riehl 21 Apr '23

21 Apr '23

Xinhao Liu will present his General Exam "Alignment Problems in Spatially Resolved Transcriptomics Data" on Wednesday, April 26, 2023 at 2:00 PM in CS 401. Committee Members: Ben Raphael (advisor), Yuri Pritykin, Olga Troyanskaya Abstract: Spatially resolved transcriptomics (SRT) technologies measure mRNA expression at thousands of locations in a tissue slice. However, nearly all SRT technologies measure expression in two dimensional slices extracted from a three-dimensional tissue, thus losing information that is shared across multiple slices from the same tissue. Integrating SRT data across multiple slices can help recover this information and improve downstream expression analyses, but multi-slice alignment and integration remains a challenging task. Existing methods for integrating SRT data either do not use spatial information or assume that the morphology of the tissue is largely preserved across slices, an assumption that is often violated due to biological or technical reasons. We introduce PASTE2, a method for partial alignment and 3D reconstruction of multi-slice SRT datasets, allowing only partial overlap between aligned slices and/or slice-specific cell types. PASTE2 formulates a novel partial Fused Gromov-Wasserstein Optimal Transport problem, which we solve using a conditional gradient algorithm. PASTE2 includes a model selection procedure to estimate the fraction of overlap between slices, and optionally uses information from histological images that accompany some SRT experiments. We show on both simulated and real data that PASTE2 obtains more accurate alignments than existing methods. We further use PASTE2 to reconstruct a 3D map of gene expression in a Drosophila embryo from a 16 slice Stereo-seq dataset. PASTE2 produces accurate alignments of multi-slice datasets from multiple SRT technologies, enabling detailed studies of spatial gene expression across a wide range of biological applications Reading List: [ https://docs.google.com/document/d/1JhXxEUxx8JpUv_wLDHTp44LG-cg0EKcLVX647y5… | https://docs.google.com/document/d/1JhXxEUxx8JpUv_wLDHTp44LG-cg0EKcLVX647y5… ] Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0

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