- Ai - Engineering Mailing Lists

AI Alignment Reading Group
by Pope, Quintin 26 May '22

26 May '22

Hello everyone, We'll be meeting at 2 PM PST on Friday to discuss the paper The Boltzmann Policy Distribution: Accounting for Systematic Suboptimality in Human Models <https://arxiv.org/abs/2204.10759>. Models of human behavior for prediction and collaboration tend to fall into two categories: ones that learn from large amounts of data via imitation learning, and ones that assume human behavior to be noisily-optimal for some reward function. The former are very useful, but only when it is possible to gather a lot of human data in the target environment and distribution. The advantage of the latter type, which includes Boltzmann rationality, is the ability to make accurate predictions in new environments without extensive data when humans are actually close to optimal. However, these models fail when humans exhibit systematic suboptimality, i.e. when their deviations from optimal behavior are not independent, but instead consistent over time. Our key insight is that systematic suboptimality can be modeled by predicting policies, which couple action choices over time, instead of trajectories. We introduce the Boltzmann policy distribution (BPD), which serves as a prior over human policies and adapts via Bayesian inference to capture systematic deviations by observing human actions during a single episode. The BPD is difficult to compute and represent because policies lie in a high-dimensional continuous space, but we leverage tools from generative and sequence models to enable efficient sampling and inference. We show that the BPD enables prediction of human behavior and human-AI collaboration equally as well as imitation learning-based human models while using far less data. 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 Pope

1 0

AI seminar: May 25, 2022
by Mangannavar, Rajesh Devaraddi 25 May '22

25 May '22

Dear all, Our next AI seminar on *"Toward Addressing Evaluation and Explanation Challenges in Scientific ML Applications" *by Shusen Liu is scheduled to be on May 25th (tomorrow), 1-2 PM PST (Add to google calendar <https://calendar.google.com/event?action=TEMPLATE&tmeid=YXExYTZuY2NiZXRhbHJ…> ). It will be followed by a 30-minute Q&A session by the graduate students. Zoom Link: https://oregonstate.zoom.us/j/93591935144?pwd=YjZaSjBYS0NmNUtjQzBEdzhPeDZ5U… *Toward Addressing Evaluation and Explanation Challenges in Scientific ML Applications* Shusen Liu Research Scientist Machine Intelligence Group Lawrence Livermore National Laboratory *Abstract:* Although the influence of deep learning in scientific domains is unmistakable, there are still fundamental barriers to utilizing these complex models for scientific discovery due in part to our inability to directly translate their predictive capabilities into scientific understanding. The root of the problem is twofold: 1) the challenge to evaluate the model in the context of the application; 2) the difficulties of reasoning about such models in terms that domain scientists can easily understand for knowledge extraction. This talk will provide a closer look at these unique challenges associated with applying deep learning to scientific applications. And cover some of our works for addressing the evaluation and explanation challenges in Scientific ML. Specifically, we will show how topological data analysis plays a crucial role in evaluating deep surrogate models for fusion science, and how deep generative models allow us to explore hypothetical materials and obtain actionable explanations that lead to improved material performance. *Speaker Bio:* Shusen Liu is a research scientist with the Machine Intelligence Group at the Lawrence Livermore National Laboratory (LLNL). His interests include fundamental research in explainable AI and high-dimensional data visualization, as well as their impact on scientific applications for advancing domain understanding. He received a Ph.D. in <http://ph.d.in/> computing from the University of Utah in 2017. *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: -> For graduate students in the AI program, attendance is strongly encouraged.

1 0

Alignment Reading Group Meeting Canceled
by Pope, Quintin 20 May '22

20 May '22

Hello everyone, I'm canceling our meeting today. I've been busy with other projects and haven't had time to prepare. If you're interested, I'd been intending to discuss The Boltzmann Policy Distribution: Accounting for Systematic Suboptimality in Human Models <https://arxiv.org/abs/2204.10759>. Models of human behavior for prediction and collaboration tend to fall into two categories: ones that learn from large amounts of data via imitation learning, and ones that assume human behavior to be noisily-optimal for some reward function. The former are very useful, but only when it is possible to gather a lot of human data in the target environment and distribution. The advantage of the latter type, which includes Boltzmann rationality, is the ability to make accurate predictions in new environments without extensive data when humans are actually close to optimal. However, these models fail when humans exhibit systematic suboptimality, i.e. when their deviations from optimal behavior are not independent, but instead consistent over time. Our key insight is that systematic suboptimality can be modeled by predicting policies, which couple action choices over time, instead of trajectories. We introduce the Boltzmann policy distribution (BPD), which serves as a prior over human policies and adapts via Bayesian inference to capture systematic deviations by observing human actions during a single episode. The BPD is difficult to compute and represent because policies lie in a high-dimensional continuous space, but we leverage tools from generative and sequence models to enable efficient sampling and inference. We show that the BPD enables prediction of human behavior and human-AI collaboration equally as well as imitation learning-based human models while using far less data. All the best, Quintin Pope

1 0

[CoRIS-List] Reminder: Today: OSU Robotics Seminar 05/20: Reid Simmons, "Agents that Learn and Teach"
by Ross L. Hatton 20 May '22

20 May '22

Hi All, As a reminder today's OSU Robotics seminar will be presented by Reid Simmons from the Carnegie Mellon University, on "Agents that Learn and Teach". The seminar will be from 10-11am, followed by an 11-11:30 student-only Q&A session. The seminar will be an *in person* event, simulcast over Zoom, and will be immediately followed by a half-hour students-only Q&A session with the speaker. May 20, 10am LINC 268 https://oregonstate.zoom.us/j/93908195212?pwd=cEdoQVdac0JxMUpOSS9xVzZjVG5xU… <https://oregonstate.zoom.us/j/93908195212?pwd=cEdoQVdac0JxMUpOSS9xVzZjVG5xU…> ----------------- Seminar Title: Agents that Learn and Teach Abstract:This talk will present recent work by my students who are investigating how agents can learn from people and teach them effectively and efficiently. In terms of learning, we have explored having agents use multiple modes of interaction, such as providing demonstrations and indicating preferences, to more effectively learn personal preferences. On the teaching side, we have explored having agents use counterfactual demonstrations, based on mental models of the learner, to efficiently scaffold learning of the agent's policies. We have also explored how providing non-verbal affective feedback can improve human learning, even if the feedback does not provide any information relevant to the task at hand. I'll also speculate on how this research can be used in assisting the elderly and their caregivers, as part of the new NSF AI-CARING Institute, of which CMU and OSU are both participants. Bio: Reid Simmons is a Research Professor in Robotics and Computer Science at Carnegie Mellon University. He is also the director of the first-in-the-country undergraduate major in Artificial Intelligence. Dr. Simmons earned his PhD from MIT in 1988 in the field of Artificial Intelligence. Since coming to CMU in 1988, he has focused on developing self-reliant robots that can autonomously operate over extended periods of time in unknown, unstructured environments. Specific research interests include human-robot social interaction, especially non-verbal communication through affect, proxemics, motion, and gesture, task planning under uncertainty, execution monitoring and failure recovery, and coordination of multiple heterogeneous robots. He also served as a Program Director at the National Science Foundation, where he oversaw the National Robotics Initiative and Smart and Autonomous Systems programs. =========== Ross L. Hatton Associate Professor, Robotics and Mechanical Engineering Collaborative Robotics and Intelligent Systems Institute Oregon State University coris.oregonstate.edu <http://coris.oregonstate.edu/> research.engr.oregonstate.edu/lram/ <http://research.engr.oregonstate.edu/lram/> rosslhatton.com <http://rosslhatton.com/> ross.hatton(a)oregonstate.edu <mailto:ross.hatton@oregonstate.edu>

1 0

AI seminar: May 18, 2022
by Mangannavar, Rajesh Devaraddi 17 May '22

17 May '22

Dear all, Our next AI seminar on *"Few-Shot Object Segmentation" *by professor Sinisa Todorovic is scheduled to be on May 18th(Tomorrow), 1-2 PM PST (Add to google calendar <https://calendar.google.com/event?action=TEMPLATE&tmeid=YXExYTZuY2NiZXRhbHJ…> ). It will be followed by a 30-minute Q&A session by the graduate students. Zoom Link: https://oregonstate.zoom.us/j/93591935144?pwd=YjZaSjBYS0NmNUtjQzBEdzhPeDZ5U… *Few-Shot Object Segmentation* Sinisa Todorovic Professor Computer Science Oregon State University *Abstract:* This talk will be about few-shot instance segmentation in images -- a basic computer vision problem arising in many applications with access to only a few examples of target object classes due to, e.g., their rarity. Our main research accomplishments in this area will be presented, including: explicit modeling and discovery of latent object parts shared across object classes; data uncertainty map estimation for each instance segmentation and data uncertainty based regularization of our few-shot learning; and model uncertainty estimation with an efficient approximation based on the probit function. Our contributions produce statistically significant performance gains over the state of the art on the benchmark COCO dataset. *Speaker Bio:* Sinisa Todorovic is a Professor in the School of Electrical Engineering and Computer Science at Oregon State University (OSU). He joined OSU in 2008, after three years of conducting postdoctoral research in the Beckman Institute at University of Illinois Urbana-Champaign (UIUC). He received his Ph.D. degree in electrical and computer engineering at University of Florida. Todorovic conducts research in computer vision with focus on object and human-activity recognition. He has served as a co-Program Chair of IEEE International Conference on Automatic Face and Gesture Recognition FG 2015, co-organized a number of tutorials and workshops on stochastic image grammars at top vision conferences, and currently serves as Editor-in-Chief of the Image and Vision Computing journal. His research has been funded by a number of federal agencies, including NSF and DARPA, as well as the IT industry. He received Jack Neubauer Best Paper Award from IEEE Transactions on Vehicular Technology in 2014, and the Oregon State College of Engineering Research Collaboration Award in 2016. This talk will be followed by a talk on the topic " *Toward Addressing Evaluation and Explanation Challenges in Scientific ML Applications*" by Shusen Liu on 25th May (Wednesday) 1-2 PM, 2022 *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: -> For graduate students in the AI program, attendance is strongly encouraged.

1 0

[CoRIS-List] OSU Robotics Seminar 05/20: Reid Simmons, "Agents that Learn and Teach"
by Ross L. Hatton 16 May '22

16 May '22

Hi All, This week's OSU Robotics seminar will be presented by Reid Simmons from the Carnegie Mellon University, on "Agents that Learn and Teach". The seminar will be from 10-11am, followed by an 11-11:30 student-only Q&A session. The seminar will be an *in person* event, simulcast over Zoom, and will be immediately followed by a half-hour students-only Q&A session with the speaker. May 20, 10am LINC 268 https://oregonstate.zoom.us/j/93908195212?pwd=cEdoQVdac0JxMUpOSS9xVzZjVG5xU… <https://oregonstate.zoom.us/j/93908195212?pwd=cEdoQVdac0JxMUpOSS9xVzZjVG5xU…> ----------------- Seminar Title: Agents that Learn and Teach Abstract:This talk will present recent work by my students who are investigating how agents can learn from people and teach them effectively and efficiently. In terms of learning, we have explored having agents use multiple modes of interaction, such as providing demonstrations and indicating preferences, to more effectively learn personal preferences. On the teaching side, we have explored having agents use counterfactual demonstrations, based on mental models of the learner, to efficiently scaffold learning of the agent's policies. We have also explored how providing non-verbal affective feedback can improve human learning, even if the feedback does not provide any information relevant to the task at hand. I'll also speculate on how this research can be used in assisting the elderly and their caregivers, as part of the new NSF AI-CARING Institute, of which CMU and OSU are both participants. Bio: Reid Simmons is a Research Professor in Robotics and Computer Science at Carnegie Mellon University. He is also the director of the first-in-the-country undergraduate major in Artificial Intelligence. Dr. Simmons earned his PhD from MIT in 1988 in the field of Artificial Intelligence. Since coming to CMU in 1988, he has focused on developing self-reliant robots that can autonomously operate over extended periods of time in unknown, unstructured environments. Specific research interests include human-robot social interaction, especially non-verbal communication through affect, proxemics, motion, and gesture, task planning under uncertainty, execution monitoring and failure recovery, and coordination of multiple heterogeneous robots. He also served as a Program Director at the National Science Foundation, where he oversaw the National Robotics Initiative and Smart and Autonomous Systems programs. =========== Ross L. Hatton Associate Professor, Robotics and Mechanical Engineering Collaborative Robotics and Intelligent Systems Institute Oregon State University coris.oregonstate.edu <http://coris.oregonstate.edu/> research.engr.oregonstate.edu/lram/ <http://research.engr.oregonstate.edu/lram/> rosslhatton.com <http://rosslhatton.com/> ross.hatton(a)oregonstate.edu

1 0

AI Alignment Reading Group
by Pope, Quintin 13 May '22

13 May '22

Hello everyone, We'll be meeting again this Friday (today) at 2 PM PST. We'll discuss "Training Language Models with Natural Language Feedback <https://arxiv.org/abs/2204.14146v2>". > Pretrained language models often do not perform tasks in ways that are in > line with our preferences, e.g., generating offensive text or factually > incorrect summaries. Recent work approaches the above issue by learning > from a simple form of human evaluation: comparisons between pairs of > model-generated task outputs. Comparison feedback conveys limited > information about human preferences per human evaluation. Here, we propose > to learn from natural language feedback, which conveys more information per > human evaluation. We learn from language feedback on model outputs using a > three-step learning algorithm. First, we condition the language model on > the initial output and feedback to generate many refinements. Second, we > choose the refinement with the highest similarity to the feedback. Third, > we finetune a language model to maximize the likelihood of the chosen > refinement given the input. In synthetic experiments, we first evaluate > whether language models accurately incorporate feedback to produce > refinements, finding that only large language models (175B parameters) do > so. Using only 100 samples of human-written feedback, our learning > algorithm finetunes a GPT-3 model to roughly human-level summarization. > It's a very interesting work that achieves some surprising and potentially useful results with a pretty simple method. If you're interested in getting a large language model to do what you want (without training an RL model of what you want), then the approach described here is a great place to look. Anyone interested should feel welcome to join! 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 Pope

1 0

AI seminar: May 18, 2022
by Mangannavar, Rajesh Devaraddi 13 May '22

13 May '22

Dear all, Our next AI seminar on *"Few-Shot Object Segmentation" *by professor Sinisa Todorovic is scheduled to be on May 18th, 1-2 PM PST (Add to google calendar <https://calendar.google.com/event?action=TEMPLATE&tmeid=YXExYTZuY2NiZXRhbHJ…> ). It will be followed by a 30-minute Q&A session by the graduate students. Zoom Link: https://oregonstate.zoom.us/j/93591935144?pwd=YjZaSjBYS0NmNUtjQzBEdzhPeDZ5U… *Few-Shot Object Segmentation* Sinisa Todorovic Computer Science Oregon State University *Abstract:* This talk will be about few-shot instance segmentation in images -- a basic computer vision problem arising in many applications with access to only a few examples of target object classes due to, e.g., their rarity. Our main research accomplishments in this area will be presented, including: explicit modeling and discovery of latent object parts shared across object classes; data uncertainty map estimation for each instance segmentation and data uncertainty based regularization of our few-shot learning; and model uncertainty estimation with an efficient approximation based on the probit function. Our contributions produce statistically significant performance gains over the state of the art on the benchmark COCO dataset. *Speaker Bio:* Sinisa Todorovic is a Professor in the School of Electrical Engineering and Computer Science at Oregon State University (OSU). He joined OSU in 2008, after three years of conducting postdoctoral research in the Beckman Institute at University of Illinois Urbana-Champaign (UIUC). He received his Ph.D. degree in electrical and computer engineering at University of Florida. Todorovic conducts research in computer vision with focus on object and human-activity recognition. He has served as a co-Program Chair of IEEE International Conference on Automatic Face and Gesture Recognition FG 2015, co-organized a number of tutorials and workshops on stochastic image grammars at top vision conferences, and currently serves as Editor-in-Chief of the Image and Vision Computing journal. His research has been funded by a number of federal agencies, including NSF and DARPA, as well as the IT industry. He received Jack Neubauer Best Paper Award from IEEE Transactions on Vehicular Technology in 2014, and the Oregon State College of Engineering Research Collaboration Award in 2016. *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: -> For graduate students in the AI program, attendance is strongly encouraged.

1 0

AI seminar: May 11, 2022
by Mangannavar, Rajesh Devaraddi 10 May '22

10 May '22

Dear all, Our next AI seminar on *"Graphical Models for Handling Data that may be Incomplete or Non-iid" *by professor Karthika Mohan and Chi Zhang is scheduled to be on May 11th (Tomorrow), 1-2 PM PST (Add to google calendar <https://calendar.google.com/event?action=TEMPLATE&tmeid=YXExYTZuY2NiZXRhbHJ…> ). It will be followed by a 30-minute Q&A session by the graduate students. Zoom Link: https://oregonstate.zoom.us/j/93591935144?pwd=YjZaSjBYS0NmNUtjQzBEdzhPeDZ5U… *Graphical Models for Handling Data that may be Incomplete or Non-iid* *Karthika Mohan *Computer Science Oregon State University * Chi Zhang* Computer Science University of California, Los Angeles *Abstract:* Majority of modern machine learning algorithms assume that data are iid and missingness/incompleteness are random. However, these assumptions rarely hold in real world datasets. For instance, the wealthy are less likely to reveal their income (income causes its own missingness) and an individual’s decision to get vaccinated is influenced by his family, friends and work colleagues. In this talk we will discuss the following, (i) modeling non-iid and missing data generation processes using causal Bayesian networks, (ii) determining whether or not a quantity of interest is recoverable under missingness, (iii) identifying conditions under which bias is to be expected given non-iid data and finally (iv) algorithm for eliminating bias given non-iid data. *Speaker Bio:* *Karthika Mohan* is an Assistant Professor of Computer Science in the School of EECS at Oregon State University. Before joining Oregon State University she was a postdoctoral scholar in the Computer Science department at University of California, Berkeley mentored by Stuart Russell. Karthika received her PhD in Computer Science (Artificial Intelligence) from University of California, Los Angeles (UCLA) where she was advised by Judea Pearl. Her research is of an interdisciplinary nature and her areas of interest include Causal Inference, Graphical Models and AI Safety. She was awarded the Google Outstanding Graduate Research Award, 2017 which is a UCLA Commencement Award. Currently she serves on the editorial board of the Journal of Causal Inference. *Chi Zhang* is a graduate student advised by Judea Pearl at the Computer Science Department in University of California, Los Angeles. Her research is mainly about causality, with a focus on causal inference under interference. *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: -> For graduate students in the AI program, attendance is strongly encouraged.

1 0

[CoRIS-List] Today: OSU Robotics Seminar 05/06: Shai Revzen, "Faster Models for Multi-Contact Motion"
by Ross L. Hatton 06 May '22

06 May '22

Hi All, As a reminder, this week's OSU Robotics seminar will be presented by Shai Revzen from the University of Michigan, on "Faster Models for Multi-Contact Motion". The seminar will be from 10-11am, followed by an 11-11:30 student-only Q&A session. The seminar will be an *in person* event, simulcast over Zoom, and will be immediately followed by a half-hour students-only Q&A session with the speaker. May 6, 10am LINC 268 https://oregonstate.zoom.us/j/93908195212?pwd=cEdoQVdac0JxMUpOSS9xVzZjVG5xU… <https://oregonstate.zoom.us/j/93908195212?pwd=cEdoQVdac0JxMUpOSS9xVzZjVG5xU…> ----------------- Seminar Title: Faster Models for Multi-Contact Motion Abstract: Multi-legged robots have many potential applications, but their motion planning presents a challenge - e.g. when slipping is involved, and when multi-contact collisions can occur. We present two recent advances that promise both a deeper insight into the physics of these systems, and large computational gains in their modeling. The first advance relates to the realization that simple "Stokesian" (first order) equations of motion govern multi-legged locomotion. Using this observation we created fast learning algorithms that identify highly predictive models for the interaction physics of multi-contact gaits from a few dozen cycles of motion. We have also produced simulation codes that are remarkably simple computationally, yet fit our experimental observations to 90% accuracy or better. The second advance relates to the resolution of multi-contact collisions. We have shown that although they are discontinuous, such collisions often admit a first order approximation that can be computed thousands of times faster than conventional collision detection methods. Combined, these methods may open up a new era of autonomous multi-legged and multi-fingered robotic systems. The work presented was funded by the NSF CMMI 1825918, ARO W911NF-14-1-0573, ARO W911NF-17-1-0306, and the D. Dan and Betty Kahn Michigan-Israel Partnership for Research and Education Autonomous Systems Mega-Project. Bio: Shai Revzen is an Associate Professor of Electrical Engineering and Computer Science in the College of Engineering of the University of Michigan, Ann Arbor. He also holds a courtesy faculty appointment in the Department of Ecology and Evolutionary Biology, and positions in the Robotics Institute and the Applied Physics Program. Currently, he is also a Visiting Associate Professor in Mechanical Engineering in the Technion, Israel. He received his PhD in Integrative Biology from Berkeley, and a B.Sc. and M.Sc. from Hebrew University in Jerusalem. He did his postdoctoral work here, in the GRASP Laboratory of the University of Pennsylvania. Shai is also CSO of Acculine, an Israeli biomedical technology start-up. In his spare time he is in the second year of law school, studying toward a JD at Wayne State University. =========== Ross L. Hatton Associate Professor, Robotics and Mechanical Engineering Collaborative Robotics and Intelligent Systems Institute Oregon State University coris.oregonstate.edu <http://coris.oregonstate.edu/> research.engr.oregonstate.edu/lram/ <http://research.engr.oregonstate.edu/lram/> rosslhatton.com <http://rosslhatton.com/> ross.hatton(a)oregonstate.edu

1 0