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

Princeton Robotics Seminar: Vikas Sindhwani, Friday June 9 at 11am
by Emily C. Lawrence 12 Sep '23

12 Sep '23

Speaker: Vikas Sindhwani, Senior Staff Research Scientist, Google Deepmind Date: Friday, June 9, 2023 Time: 11am ET Location: Virtual, Zoom link: [ https://princeton.zoom.us/my/robotics | https://princeton.zoom.us/my/robotics ] Event page: [ https://robo.princeton.edu/seminar | https://robo.princeton.edu/seminar ] Title: Large Language Models with Eyes, Arms and Legs Abstract: To become useful in human-centric environments, robots must demonstrate language comprehension, semantic understanding and logical reasoning capabilities working in concert with low-level physical skills. With the advent of modern "foundation models" trained on massive datasets, the algorithmic path to developing general-purpose “robot brains” is (arguably) becoming clearer, though many challenges remain. In the first part of this talk, I will attempt to give a flavor of how state-of-the-art multimodal foundation models are built, and how they can be bridged with low-level control. In the second part of the talk, I will summarize a few surprising lessons on control synthesis observed while solving a collection of Robotics benchmarks at Google. I will end with some emerging open problems and opportunities at the intersection of dynamics, control and foundation models. Bio: Vikas Sindhwani is Research Scientist at Google Deepmind in New York where he leads a research group focused on solving a range of planning, perception, learning, and control problems arising in Robotics. His interests are broadly in core mathematical foundations of statistical machine learning, and in end-to-end design aspects of building large-scale and robust AI systems. He received the best paper award at Uncertainty in Artificial Intelligence (UAI-2013), the IBM Pat Goldberg Memorial Award in 2014, and was finalist for Outstanding Planning Paper Award at ICRA-2022. He serves on the editorial board of Transactions on Machine Learning Research (TMLR) and IEEE Transactions on Pattern Analysis and Machine Intelligence; he has been area chair and senior program committee member for NeurIPS, International Conference on Learning Representations (ICLR) and Knowledge Discovery and Data Mining (KDD). He previously headed the Machine Learning group at IBM Research, NY. He has a PhD in Computer Science from the University of Chicago and a B.Tech in Engineering Physics from Indian Institute of Technology (IIT) Mumbai. His publications are available at: http://vikas.sindhwani.org/

1 5

Aquinas Hobor (University College London) July 28 at 10:30
by Michael Estepp 25 Jul '23

25 Jul '23

Friday, July 28, 10:30am in room CS105: Programming Languages Seminar Title: Smart Learning to Find Dumb Contracts (by Abdelaziz & Hobor, to appear in USENIX Security 2023) Speaker: Aquinas Hobor (University College London) Abstract: We introduce the Deep Learning Vulnerability Analyzer (DLVA), a vulnerability detection tool for Ethereum smart contracts based on powerful deep learning techniques for sequential data adapted for bytecode. We train DLVA to judge bytecode even though the supervising oracle, Slither, can only judge source code. DLVA's training algorithm is general: we "extend" a source code analysis to bytecode without any manual feature engineering, predefined patterns, or expert rules. DLVA's training algorithm is also robust: it overcame a 1.25% error rate mislabelled contracts, and-the student surpassing the teacher-found vulnerable contracts that Slither mislabelled. In addition to extending a source code analyser to bytecode, DLVA is much faster than conventional tools for smart contract vulnerability detection based on formal methods: DLVA checks contracts for 29 vulnerabilities in 0.2 seconds, a 10-1,000x speedup compared to traditional tools. We benchmark DLVA against nine well-known smart contract analysis tools. Despite using much less analysis time, DLVA completed every query, leading the pack with an average accuracy of 99.7%, pleasingly balancing high true positive rates with low false positive rates.

1 0

Qinshi Wang will present his FPO "Foundationally Verified Data Plane Programming" on Thursday, August 24, 2023 at 2:30pm in CS 402
by Nicki Mahler 25 Jul '23

25 Jul '23

Qinshi Wang will present his FPO "Foundationally Verified Data Plane Programming" on Thursday, August 24, 2023 at 2:30pm in CS 402. The members of his committee are as follows: Readers: David Walker, Nate Foster (Cornell) Examiners: Andrew Appel (Adviser), Jennifer Rexford, Zachary Kincaid All are welcome to attend. Abstract: P4 is a major standardized programming language for programming and specifying the network data plane. P4 is widely applied in a variety of network functionalities, including monitoring, traffic management, forwarding, and security. Recently, stateful applications have been emerging in this area, as supported by programmable hardware. Typical stateful applications include network telemetry (heavy hitters, distributed denial-of-service (DDoS) detection, performance monitoring), middleboxes (firewalls, network address translation (NAT), load balancers, intrusion detection), and distributed services (in-network caching, lock management, conflict detection). Their complexity and rich properties are beyond the ability of existing P4 verifiers. In this thesis, we propose Verifiable P4: a new framework for P4 program verification based on interactive theorem proving that is (1) capable of proving multi-packet properties, (2) modular in terms of P4 program modules, and (3) foundationally sound with respect to a mechanized formal semantics of P4. In order to achieve these goals, we built (1) a mechanized formal semantics of P4 more comprehensive and convenient than existing formal semantics, (2) a set of program logic rules that are proven sound, and (3) an interactive verification system based on the program logic and Coq tactic mechanism. We verified a stateful firewall fully implemented in P4 using the sliding-window Bloom filter with Verifiable P4 and evaluated its utility.

1 0

Zhenyu Song will present his FPO "Optimizing Content Distribution Network Caches with Machine Learning" on Tuesday, August 1, 2023 at 1pm in CS 402.
by Gradinfo 19 Jul '23

19 Jul '23

Zhenyu Song will present his FPO "Optimizing Content Distribution Network Caches with Machine Learning" on Tuesday, August 1, 2023 at 1pm in CS 402. The members of his committee are as follows: Readers: Ravi Netravali Daniel S. Berger (Microsoft & UW) Examiners: Kai Li (Adviser), Wyatt Lloyd (Adviser), and Ryan Adams All are welcome to attend. Abstract: Content Distribution Networks (CDNs) play a pivotal role in Internet traffic. A key part of this caching mechanism is the eviction algorithm that handles the replacement of old cached objects. The effectiveness of the eviction algorithm significantly influences CDN performance. This dissertation explores the application of machine learning (ML) to optimize cache eviction algorithms in CDNs. The central questions addressed in this work are: how to utilize ML to devise an eviction algorithm that surpasses existing heuristics on byte miss ratio, and how to mitigate the CPU overhead while enhancing the robustness of a learned cache in large-scale deployment. Two major challenges faced in the design of a learning-based cache eviction algorithm include heterogeneous user access patterns across different locations and times, and computational and space overheads. To address these challenges, we developed two ML-based eviction algorithms, Learning Relaxed Belady (LRB) and Heuristic Aided Learned Preference (HALP). LRB is the first CDN cache algorithm to directly approximate the Belady MIN (oracle) algorithm by learning access patterns, providing a significant improvement over traditional eviction algorithms. It demonstrated a WAN traffic reduction of 4-25% across six production CDN traces in our simulation. HALP, on the other hand, achieves low CPU overhead and robust DRAM byte miss ratio improvement by augmenting a heuristic policy with ML. It has shown to reduce DRAM byte miss during peak by an average of 9.1%, with a modest CPU overhead of 1.8%, while deployed in YouTube CDN production clusters.

1 0