News

December 6-14, 2021 | NeurIPS Proceedings

(Works acknowledging TILOS support are denoted by **)

Risk Bounds for Over-parameterized Maximum Margin Classification on Sub-Gaussian Mixtures Yuan Cao, Quanquan Gu, Mikhail Belkin
Multiple Descent: Design Your Own Generalization Curve Lin Chen, Yifei Min, Mikhail Belkin, Amin Karbasi
NN-Baker: A Neural-network Infused Algorithmic Framework for Optimization Problems on Geometric Intersection Graphs Evan McCarty, Qi Zhao, Anastasios Sidiropoulos, Yusu Wang
Fair Classification with Adversarial Perturbations L. Elisa Celis, Anay Mehrotra, Nisheeth Vishnoi
Coresets for Time Series Clustering Lingxiao Huang, K Sudhir, Nisheeth Vishnoi
Support Recovery of Sparse Signals from a Mixture of Linear Measurements Soumyabrata Pal, Arya Mazumdar, Venkata Gandikota
Fuzzy Clustering with Similarity Queries Wasim Huleihel, Arya Mazumdar, Soumyabrata Pal
L2ight: Enabling On-Chip Learning for Optical Neural Networks via Efficient in-situ Subspace Optimization Jiaqi Gu, Hanqing Zhu, Chenghao Feng, Zixuan Jiang, Ray Chen, David Pan
Reducing Collision Checking for Sampling-Based Motion Planning Using Graph Neural Networks Chenning Yu, Sicun Gao
Adaptive Sampling for Minimax Fair Classification Shubhanshu Shekhar, Greg Fields, Mohammad Ghavamzadeh, Tara Javidi
NovelD: A Simple yet Effective Exploration Criterion Tianjun Zhang, Huazhe Xu, Xiaolong Wang, Yi Wu, Kurt Keutzer, Joseph E. Gonzalez, Yuandong Tian
Multi-Person 3D Motion Prediction with Multi-Range Transformers Jiashun Wang, Huazhe Xu, Medhini Narasimhan, Xiaolong Wang **
Stabilizing Deep Q-Learning with ConvNets and Vision Transformers under Data Augmentation Nicklas Hansen, Hao Su, Xiaolong Wang **
An Exponential Improvement on the Memorization Capacity of Deep Threshold Networks Shashank Rajput, Kartik Sreenivasan, Dimitris Papailiopoulos, Amin Karbasi
Submodular + Concave Siddharth Mitra, Moran Feldman, Amin Karbasi
Parallelizing Thompson Sampling Amin Karbasi, Vahab Mirrokni, Mohammad Shadravan
Model-Based Domain Generalization Alexander Robey, George Pappas, Hamed Hassani
Linear Convergence in Federated Learning: Tackling Client Heterogeneity and Sparse Gradients Aritra Mitra, Rayana Jaafar, George Pappas, Hamed Hassani
Adversarial Robustness with Semi-Infinite Constrained Learning Alexander Robey, Luiz Chamon, George Pappas, Hamed Hassani, Alejandro Ribeiro
Three Operator Splitting with Subgradients, Stochastic Gradients, and Adaptive Learning Rates Alp Yurtsever, Alex Gu, Suvrit Sra
Scaling up Continuous-Time Markov Chains Helps Resolve Underspecification Alkis Gotovos, Rebekka Burkholz, John Quackenbush, Stefanie Jegelka
Measuring Generalization with Optimal Transport Ching-Yao Chuang, Youssef Mroueh, Kristjan Greenewald, Antonio Torralba, Stefanie Jegelka
Can contrastive learning avoid shortcut solutions? Joshua Robinson, Li Sun, Ke Yu, Kayhan Batmanghelich, Stefanie Jegelka, Suvrit Sra
What training reveals about neural network complexity Andreas Loukas, Marinos Poiitis, Stefanie Jegelka

October 11, 2021 | Quanta Magazine

(Anil Ananthaswamy's article in Quanta Magazine discusses kernel methods and their relevance, including some of Misha Belkin's recent work.)

In the machine learning world, the sizes of artificial neural networks and their outsize successes are creating conceptual conundrums. When a network named Alexnet won an annual image recognition competition in 2021, it had about 60 million parameters. These parameters, fine-tuned during training, allowed AlexNet to recognize images that it had never seen before. Two years later, a network named VGG wowed the competition with more than 130 million such parameters. Some artificial neural networks or ANNs, now have billions of parameters.

Full article here.

August 2, 2021 | National University

LA JOLLA, CALIF — As part of an initiative to establish a series of artificial intelligence research institutes nationwide, the National Science Foundation (NSF) has awarded a $20 million grant to a partnership of prestigious universities led by University of California, San Diego, and including National University, Yale University, the Massachusetts Institute of Technology, the University of Pennsylvania, and the University of Texas at Austin.

Full article here.

July 29, 2021 | University of Texas at Austin

The National Science Foundation just announced 11 new artificial intelligence institutes across the nation, and researchers from the Cockrell School of Engineering at The University of Texas at Austin will play prominent roles in two of them.

The investment from NSF, $220 million in total, underscores the importance of AI in research today. The goals of these institutes include helping older adults lead more independent lives and improving the quality of their care; transforming AI into a more accessible “plug-and-play” technology; creating solutions to improve agriculture and food supply chains; enhancing adult online learning by introducing AI as a foundational element; and supporting underrepresented students in elementary to postdoctoral STEM education to improve equity and representation in AI research.

Full article here.

July 29, 2021 | MIT Computer Science & Artificial Intelligence Lab

The National Science Foundation (NSF) announced today an investment of $220 million to establish 11 artificial intelligence (AI) institutes, each receiving $20 million over five years. One of these, The Institute for Learning-enabled Optimization at Scale (TILOS), will be led by the University of California San Diego in partnership with the Massachusetts Institute of Technology; San Diego-based National University; the University of Pennsylvania; the University of Texas atAustin; and Yale University. TILOS is also partially supported by Intel Corporation.

Full article here.

July 29, 2021 | Source: Yale School of Engineering & Applied Science

Yale researchers are taking a major role in two artificial intelligence (AI) institutes, each funded by the National Science Foundation (NSF).

The NSF announced today that it is investing $220 million to establish 11 AI institutes, each receiving $20 million over five years. Yale researchers are taking leadership positions in two of these institutes, one of which will focus on optimization, and another dedicated to edge computing and computer network systems. Eight Yale scientists in total are members of the institutes.

Full article here.

July 29, 2021 | Source: Newswise

The National Science Foundation (NSF) announced today an investment of $220 million to establish 11 artificial intelligence (AI) institutes, each receiving $20 million over five years. One of these, The Institute for Learning-enabled Optimization at Scale (TILOS), will be led by the University of California San Diego in partnership with the Massachusetts Institute of Technology; San Diego-based National University; the University of Pennsylvania; the University of Texas at Austin; and Yale University. TILOS is also partially supported by Intel Corporation.

Full article here.

July 29, 2021 | Source: Jacobs School of Engineering News

SAN DIEGO — The National Science Foundation (NSF) announced today an investment of $220 million to establish 11 artificial intelligence (AI) institutes, each receiving $20 million over five years. One of these, The Institute for Learning-enabled Optimization at Scale (TILOS), will be led by the University of California San Diego in partnership with the Massachusetts Institute of Technology; San Diego-based National University; the University of Pennsylvania; the University of Texas at Austin; and Yale University. TILOS is also partially supported by Intel Corporation.

Full article here.

Proposer’s Name	Corresponding Email	Project Title	Brief Description (what is sought)	List of Existing Collaborators (one per line)	List of References	Additional Information	Created
Andrew Kahng	abk-tilos@eng.ucsd.edu	Identifying "doomed runs" of an IC routing tool	The “detailed routing” step of IC place-and-route can take many hours or days. The routing tool must construct a set of trees of wire segments, where each tree connects a prescribed set of terminals of logic gates in the chip design. The trees cannot touch each other. Routing tools perform many iterations of “ripup-and-reroute” to gradually reduce the number of design rule violations in the routing solution. But if the tool cannot reduce the violations to a usable level (e.g., 2000 violations can be fixed by an expert engineer in a week), then the run is useless. This project will try to predict “doomed runs” (or, a final number of violations) based on information available at the initial iterations of the routing tool. Being able to terminate “doomed runs”, without terminating runs that will eventually succeed, has very high value in the chip design process.	Dr. Seungwon Kim, sek006@eng.ucsd.edu	(1) Slides 10, 11, 12 in: https://vlsicad.ucsd.edu/NEWS18/malenda-kahng-talk-distributed.pdf (workshop talk at DAC-2018) (2) GitHub: TritonRoute https://github.com/The-OpenROAD-Project/OpenROAD/tree/master/src/drt (3) A. B. Kahng, L. Wang and B. Xu, "TritonRoute-WXL: The Open Source Router with Integrated DRC Engine", IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems (2021), doi:10.1109/TCAD.2021.3079268 https://vlsicad.ucsd.edu/Publications/Journals/j136.pdf	Google Doc open for comments (please use @abk-tilos@eng.ucsd.edu to flag for attention): https://docs.google.com/document/d/1rOHx6fYesH-XgKHKyNZcWy30lF33WvAJEEJKFBwZBj0/edit?usp=sharing	03/09/2022
Tara Javidi	tjavidi@eng.ucsd.edu	Empirical Optimization of Band-limitted (Graph) Signals For Networks	We have developed an adaptive sampling method for empirical (zeroth order) optimization of a function in reproducing kernel Hilbert space with a known kernel. We are interested in extending this work in the following directions: 1. A special case of RKHS is that of band-limited functions where our algorithm can be precisely compared to a Shannon/Nyquest sampler. This view suggests similar correspondence might be of interest when the input space consists of graphs. 2. Practically, data needs to be used to identify the kernel and estimate parameters. I am interested to work with networking and chip use case folks to see how such estimation can be optimized.	Tara Javidi Farinaz Koushanfar	Multi-Scale Zero-Order Optimization of Smooth Functions in an RKHS. https://arxiv.org/abs/2005.04832 AdaNS: Adaptive Non-Uniform Sampling for Automated Design of Compact DNNs		03/11/2022
Andrew Kahng	abk-tilos@eng.ucsd.edu	Learning Good Hierarchical Clustering in Clock Distribution	The clock signal is the “heartbeat” of any synchronous digital circuit. The clock signal is distributed from a “clock source” to every “clock sink” (flip-flop) via a rooted “clock tree” of wires and buffers. For figures, see the reference Doc. (1) The clock signal should reach every sink at roughly the same time: this is the goal of “zero skew” or “bounded skew” (e.g., Charikar et al., SODA99). (2) Also, the clock tree should have minimum total wirelength, since more wirelength means more power consumption. (3) And, the clock signal must be “refreshed” every so often, because a logic gate can only drive a limited amount of capacitive load. Thus, the clock tree has a hierarchy of “clock buffers”, each of which sends the signal onward to a bounded amount of wiring and child buffers/sinks. The goal is to learn how to hierarchically cluster sinks and buffers in clock trees, based on “good example solutions” obtained by running clock tree synthesis (CTS) in commercial place-and-route tools. Benchmark testcases and solution quality evaluators, and potentially relevant metadata/features, can be provided. See the “2013_CTS_study” link in the reference Doc.	Ravi Varadarajan, rvaradarajan@ucsd.edu Matt Liberty, mliberty@eng.ucsd.edu	M. Charikar, J. Kleinberg, R. Kumar, S. Rajagopalan, A. Sahai and A. Tomkins, “Minimizing Wirelength in Zero and Bounded Skew Clock Trees”, Proc. SODA, 1999, pp. 177-184. http://web.cs.ucla.edu/~sahai/work/web/1999Publications/SODA1999.pdf A. Z. Zelikovsky and I. I. Mandoiu, “Practical Approximation Algorithms for Zero- and Bounded-Skew Trees”, https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.38.1047&rep=rep1&type=pdf See images in HTML files linked from https://vlsicad.ucsd.edu/2013_CTS_study/index.html . E.g., “Click to change level” in https://vlsicad.ucsd.edu/2013_CTS_study/uni_high_ar2.html . Can we learn this clustering, and what vertex or layout features determine this clustering?	Google Doc open for comments (please use @abk-tilos@eng.ucsd.edu to flag for attention): https://docs.google.com/document/d/1_AQNf6N75fz6cBtoQSX3hJsQNJV20DHmTIrZO9NkCvQ/edit?usp=sharing	03/13/2022
Andrew Kahng	abk-tilos@eng.ucsd.edu	Clusters That Stay Together	A chip design can be viewed as a hypergraph with lots of metadata/labels on vertices and hyperedges. A complex sequence of heuristic algorithms will place this hypergraph onto the layout region of the chip. The goal is to reduce problem size without losing solution quality by clustering the hypergraph, such that each cluster “stays together” in the placement. I.e., we want to learn how to cluster “with no mistakes”. Even a 10X problem size reduction “with no mistakes” would have a huge impact on chip design practice. See Slide 6 of the first reference. See also the “Tutorial on chips + ML” (second reference). The existing collaborators can provide relevant data such as hypergraphs, various metadata on vertices and hyperedges, information from placements returned by a commercial place-and-route tool, etc.	Zhiang Wang, zhw033@ucsd.edu Bodhi Pramanik, bpramanik@eng.ucsd.edu	Slide 6 (see talknotes as well) in this PPTX file from the TILOS Retreat (January 6): https://docs.google.com/presentation/d/17ENPs9KLt4MoHtTRdhaXapNweC6auRlm/edit?usp=sharing&ouid=107509409457975522044&rtpof=true&sd=true Work in Progress: “Tutorial on chips + ML”, https://docs.google.com/document/d/15cr32aB1Q7kawFwNq__DvL0_CkzimXr8BTx4Xk0lFGY/edit?usp=sharing M. Fogaça, A. B. Kahng, E. Monteiro, R. Reis, L. Wang and M. Woo, "On the Superiority of Modularity-Based Clustering for Determining Placement-Relevant Clusters", (.pdf) Integration: The VLSI Journal 74 (2020), pp. 32-44.	Google Doc open for comments (please use @abk-tilos@eng.ucsd.edu to flag for attention): https://docs.google.com/document/d/12RqI4FVKcLNfaZq1wCiac3sm3C-duj3qxHYSvuWSnyE/edit?usp=sharing	03/13/2022
Justin McCarty	justinmccarty@mailerking.xyz	Ubjn	Hello, Wondering if you accept link inserts on existing posts on tilos.ai? How much would you charge for this? Also, do you have any other sites for link inserts? Justin If you'd like to unsubscribe click the following link. https://mailerking.xyz/optout/?site=tilos.ai	Hello, Wondering if you accept link inserts on existing posts on tilos.ai? How much would you charge for this? Also, do you have any other sites for link inserts? Justin If you'd like to unsubscribe click the following link. https://mailerking.xyz/optout/?site=tilos.ai	Hello, Wondering if you accept link inserts on existing posts on tilos.ai? How much would you charge for this? Also, do you have any other sites for link inserts? Justin If you'd like to unsubscribe click the following link. https://mailerking.xyz/optout/?site=tilos.ai	Hello, Wondering if you accept link inserts on existing posts on tilos.ai? How much would you charge for this? Also, do you have any other sites for link inserts? Justin If you'd like to unsubscribe click the following link. https://mailerking.xyz/optout/?site=tilos.ai	12/09/2022

TILOS in the NEWS

December 07, 2022 | AATRN

November 17, 2022 | UCSD TODAY

November, 2022 | NetSci

September 22, 2022 | BREAKTHROUGH PRIZE

August 3, 2022 | National University

February 17, 2022 | Intel Labs

January 19, 2022 | ACM Media Center

December 23, 2021 | ACM News Release

December 6-14, 2021 | NeurIPS Proceedings

October 11, 2021 | Quanta Magazine

August 27, 2021 | UC San Diego News Center

August 2, 2021 | National University

July 29, 2021 | San Diego Union Tribune

July 29, 2021 | KPBS Science and Technology

July 29, 2021 | Penn Engineering Today

July 29, 2021 | University of Texas at Austin

July 29, 2021 | MIT Computer Science & Artificial Intelligence Lab

July 29, 2021 | Source: Yale School of Engineering & Applied Science

July 29, 2021 | Source: Newswise

July 29, 2021 | Source: Intel

The Institute will pursue foundational breakthroughs at the nexus of artificial intelligence and optimization to transform chip design, robotics, and communication networks.

July 29, 2021 | Source: Jacobs School of Engineering News

July 29, 2021 | Source: NSF AI Institutes Team