December 2021 - Ai - Engineering Mailing Lists

AI Alignment Reading Group
by Pope, Quintin 14 Dec '21

14 Dec '21

Hello everyone, We will meet on Friday at 2 PM PST. We're continuing to explore transformer interpretability. Our next paper is "Knowledge Neurons in Pretrained Transformers <https://arxiv.org/abs/2104.08696>" (GitHub implementation <https://github.com/EleutherAI/knowledge-neurons>). This paper is able to identify where pretrained transformers store knowledge and suppress, amplify or modify that knowledge. If you're at all interested in how transformers learn and represent information, you're welcome to attend! Abstract: > Large-scale pretrained language models are surprisingly good at recalling > factual knowledge presented in the training corpus. In this paper, we > explore how implicit knowledge is stored in pretrained Transformers by > introducing the concept of knowledge neurons. Given a relational fact, we > propose a knowledge attribution method to identify the neurons that express > the fact. We present that the activation of such knowledge neurons is > highly correlated to the expression of their corresponding facts. In > addition, even without fine-tuning, we can leverage knowledge neurons to > explicitly edit (such as update, and erase) specific factual knowledge for > pretrained Transformers. > Join Zoom Meeting https://oregonstate.zoom.us/j/95843260079?pwd=TzZTN0xPaFZrazRGTElud0J1cnJLU… Password: 961594 Phone Dial-In Information +1 971 247 1195 US (Portland) +1 253 215 8782 US (Tacoma) +1 301 715 8592 US (Washington DC) Meeting ID: 958 4326 0079 All the best, Quintin

1 0

[CoRIS-List] Course registration (W'22): AI539 ST/ Safe and Reliable Autonomy
by Saisubramanian, Sandhya 06 Dec '21

06 Dec '21

Hi Everyone, I am offering a seminar-style special topics course AI 539: Safe and Reliable Autonomy in Winter 2022. While this course currently has AI/CS restriction on catalog, it is open to all CS/AI/Robotics students interested in safe and reliable operation of autonomous systems. Robotics students interested in this course can email me for overrides. More details about the course are available at https://www.sandhyasai.com/ai-539-st-safe-and-reliable-autonomy. Email me if you have questions about the course. Best, Sandhya Saisubramanian Assistant Professor School of EECS Oregon State University

1 0

Alignment Reading Group
by Pope, Quintin 06 Dec '21

06 Dec '21

Hello everyone, The alignment reading group will again meet on Friday at 2 PM PST. We're continuing our exploration of interpretability, this time focusing specifically on transformers. We'll discuss Transformer Feed-Forward Layers Are Key-Value Memories <https://arxiv.org/abs/2012.14913v2>. If you've ever asked yourself, "Okay, but what do transformers *actually* do?", this is a great paper to read. Newcomers are welcome! Paper abstract: > Feed-forward layers constitute two-thirds of a transformer model's > parameters, yet their role in the network remains under-explored. We show > that feed-forward layers in transformer-based language models operate as > key-value memories, where each key correlates with textual patterns in the > training examples, and each value induces a distribution over the output > vocabulary. Our experiments show that the learned patterns are > human-interpretable, and that lower layers tend to capture shallow > patterns, while upper layers learn more semantic ones. The values > complement the keys' input patterns by inducing output distributions that > concentrate probability mass on tokens likely to appear immediately after > each pattern, particularly in the upper layers. Finally, we demonstrate > that the output of a feed-forward layer is a composition of its memories, > which is subsequently refined throughout the model's layers via residual > connections to produce the final output distribution. > Join Zoom Meeting https://oregonstate.zoom.us/j/95843260079?pwd=TzZTN0xPaFZrazRGTElud0J1cnJLU… Password: 961594 Phone Dial-In Information +1 971 247 1195 US (Portland) +1 253 215 8782 US (Tacoma) +1 301 715 8592 US (Washington DC) Meeting ID: 958 4326 0079 All the best, Quintin

1 0

Alignment Reading Group: Interpretability
by Pope, Quintin 01 Dec '21

01 Dec '21

Hello everyone, The alignment reading group will meet this Friday at 2 PM PST. We’ll be discussing "Interpretability in ML: A Broad Overview <https://www.lesswrong.com/posts/57fTWCpsAyjeAimTp/interpretability-in-ml-a-…> " The article introduces many different aspects of interpretability, such as why we might want interpretable models, how to define or think about interpretability, different techniques used for different models, and how to evaluate the quality of interpretability techniques. Please join if you're interested in ML interpretability. Newcomers are welcome! Join Zoom Meeting https://oregonstate.zoom.us/j/95843260079?pwd=TzZTN0xPaFZrazRGTElud0J1cnJLU… Password: 961594 Phone Dial-In Information +1 971 247 1195 US (Portland) +1 253 215 8782 US (Tacoma) +1 301 715 8592 US (Washington DC) Meeting ID: 958 4326 0079 All the best, Quintin

1 0

AI seminar: Dec 1 2021
by Mangannavar, Rajesh Devaraddi 01 Dec '21

01 Dec '21

Dear all, Our next AI seminar on *"**AI for Imperfect-Information Games: Poker and Beyond" *by Noam Brown is scheduled to be on December 1st(Tomorrow), 1-2 PM PST. It will be followed by a 30 minute Q&A session by the graduate students. Note that this is a zoom event. Zoom Link: https://oregonstate.zoom.us/j/93591935144?pwd=YjZaSjBYS0NmNUtjQzBEdzhPeDZ5U… *Title: AI for Imperfect-Information Games: Poker and Beyond* Noam Brown Research Scientist Facebook AI Research *Abstract*: The field of artificial intelligence has had a number of high-profile successes in the domain of perfect-information games like chess and Go where all participants know the exact state of the world. But real-world strategic interactions typically involve hidden information, such as in negotiations, cybersecurity, and financial markets. Past AI techniques fall apart in these settings, with poker serving as the classic benchmark and challenge problem. In this talk I will cover the key breakthroughs behind Libratus and Pluribus, the first AI agents to defeat elite human professionals in two-player no-limit poker and multiplayer no-limit poker, respectively. In particular, I will discuss new equilibrium-finding algorithms and depth-limited search techniques for imperfect-information games that are orders of magnitude faster than prior approaches. Finally, I will conclude with recent work generalizing these results to other imperfect-information multi-agent settings beyond poker. *Bio*: Noam Brown is a Research Scientist at Facebook AI Research working on multi-agent artificial intelligence and computational game theory, with a particular focus on sequential imperfect-information games. He co-created Libratus and Pluribus, the first AIs to defeat top humans in two-player no-limit poker and multiplayer no-limit poker, respectively. He has received the Marvin Minsky Medal for Outstanding Achievements in AI, was named one of MIT Tech Review's 35 Innovators Under 35, and his work on Pluribus was named by Science Magazine to be one of the top 10 scientific breakthroughs of 2019. Noam received his PhD from Carnegie Mellon University, for which he received the IFAAMAS Victor Lesser Distinguished Dissertation Award and the CMU School of Computer Science Distinguished Dissertation Award. *Please watch this space for future AI Seminars :* * https://eecs.oregonstate.edu/ai-events <https://eecs.oregonstate.edu/ai-events>* Rajesh Mangannavar, Graduate Student Oregon State University ---- AI Seminar Important Reminders: -> The AI Seminar has a strict "no electronics" and "no recordings" policy. -> For graduate students in the AI program, attendance is strongly encouraged.

1 0