diff --git a/_layouts/organizers.html b/_layouts/organizers.html index 01d1b3e..d2b59ef 100755 --- a/_layouts/organizers.html +++ b/_layouts/organizers.html @@ -2,55 +2,57 @@ layout: default --- +
-
-

{{ page.title | escape }}

-
- -
- - {% for organizer in site.organizers %} - - - - - {% endfor %} -
-
-

- {% if organizer.img %} - - {% else %} - - {% endif %} -

-
-
-
- - {% if organizer.webpage %} - {{ organizer.name }} - {% else %} - {{ organizer.name }} - {% endif %} -
- - {% if organizer.affil_link %} - {{ organizer.affil }} - {% else %} - {{ organizer.affil }} - {% endif %} - - {% if organizer.affil2_link %} -
- {{ organizer.affil2 }} - {% elsif organizer.affil2 %} -
- {{ organizer.affil2 }} - {% endif %} -
-
-
- -
+
"If I have seen further, it is by standing on the shoulders of giants."
+ - Isaac Newton, 1675
+ +
+ + + {% for organizer in site.organizers %} + + {% endfor %} + +
+
+

+ {% if organizer.img_url %} + + {% elsif organizer.img %} + + {% else %} + + {% endif %} +

+
+
+ + + {% if organizer.webpage %} + {{ organizer.name }} + {% else %} + {{ organizer.name }} + {% endif %} + +
+ + {% if organizer.affil_link %} + {{ organizer.affil }} + {% else %} + {{ organizer.affil }} + {% endif %} + + {% if organizer.affil2_link %} +
+ {{ organizer.affil2 }} + {% elsif organizer.affil2 %} +
+ {{ organizer.affil2 }} + {% endif %} +
+
+
+ + \ No newline at end of file diff --git a/_organizers/organizerjanedoe.md b/_organizers/0_ziyiyang.md similarity index 50% rename from _organizers/organizerjanedoe.md rename to _organizers/0_ziyiyang.md index a91fbd2..287fac3 100644 --- a/_organizers/organizerjanedoe.md +++ b/_organizers/0_ziyiyang.md @@ -1,20 +1,20 @@ --- # Name of the organizer -name: Organizer Jane Doe +name: Ziyi Yang # Link to the organizer's webpage -webpage: https://jane.doe +webpage: http://yzylmc.com # Primary affiliation -affil: Buzz University +affil: 2024(Now) # Link to the primary affiliation -affil_link: https://buzz.edu +# affil_link: https://buzz.edu # An image of the organizer (square aspect ratio works the best) (place in the `assets/img/organizers` directory) -img: jane.jpg +img: ziyiyang.jpeg -# Secondary affiliation -affil2: BuzzFizz Corp -# Link to the secondary affiliation -affil2_link: https://buzzfizz.corp +# # Secondary affiliation +# affil2: BuzzFizz Corp +# # Link to the secondary affiliation +# affil2_link: https://buzzfizz.corp --- diff --git a/_organizers/1_benjaminspiegel.md b/_organizers/1_benjaminspiegel.md new file mode 100644 index 0000000..90e9092 --- /dev/null +++ b/_organizers/1_benjaminspiegel.md @@ -0,0 +1,20 @@ +--- +# Name of the organizer +name: Benjamin A. Spiegel + +# Link to the organizer's webpage +webpage: https://benjaminaspiegel.com/ + +# Primary affiliation +affil: 2023 +# Link to the primary affiliation +# affil_link: https://buzz.edu + +# An image of the organizer (square aspect ratio works the best) (place in the `assets/img/organizers` directory) +img: benjaminspiegel.png + +# # Secondary affiliation +# affil2: BuzzFizz Corp +# # Link to the secondary affiliation +# affil2_link: https://buzzfizz.corp +--- diff --git a/_organizers/2_sakettiwari.md b/_organizers/2_sakettiwari.md new file mode 100644 index 0000000..956174d --- /dev/null +++ b/_organizers/2_sakettiwari.md @@ -0,0 +1,20 @@ +--- +# Name of the organizer +name: Saket Tiwari + +# Link to the organizer's webpage +webpage: https://saketirl.github.io/ + +# Primary affiliation +affil: 2022 +# Link to the primary affiliation +# affil_link: https://buzz.edu + +# An image of the organizer (square aspect ratio works the best) (place in the `assets/img/organizers` directory) +img: sakettiwari.jpeg + +# # Secondary affiliation +# affil2: BuzzFizz Corp +# # Link to the secondary affiliation +# affil2_link: https://buzzfizz.corp +--- diff --git a/_organizers/3_kaiyuzheng.md b/_organizers/3_kaiyuzheng.md new file mode 100644 index 0000000..9819402 --- /dev/null +++ b/_organizers/3_kaiyuzheng.md @@ -0,0 +1,20 @@ +--- +# Name of the organizer +name: Kaiyu Zheng + +# Link to the organizer's webpage +webpage: https://kaiyuzheng.me/ + +# Primary affiliation +affil: 2021 +# Link to the primary affiliation +# affil_link: https://buzz.edu + +# An image of the organizer (square aspect ratio works the best) (place in the `assets/img/organizers` directory) +img: kaiyuzheng.png + +# # Secondary affiliation +# affil2: BuzzFizz Corp +# # Link to the secondary affiliation +# affil2_link: https://buzzfizz.corp +--- diff --git a/_organizers/4_samlobel.md b/_organizers/4_samlobel.md new file mode 100644 index 0000000..b7e969b --- /dev/null +++ b/_organizers/4_samlobel.md @@ -0,0 +1,20 @@ +--- +# Name of the organizer +name: Sam Lobel + +# Link to the organizer's webpage +webpage: https://samlobel.github.io/ + +# Primary affiliation +affil: 2020 +# Link to the primary affiliation +# affil_link: https://buzz.edu + +# An image of the organizer (square aspect ratio works the best) (place in the `assets/img/organizers` directory) +img: samlobel.jpeg + +# # Secondary affiliation +# affil2: BuzzFizz Corp +# # Link to the secondary affiliation +# affil2_link: https://buzzfizz.corp +--- diff --git a/_organizers/5_lucaslehnert.md b/_organizers/5_lucaslehnert.md new file mode 100644 index 0000000..14dfec4 --- /dev/null +++ b/_organizers/5_lucaslehnert.md @@ -0,0 +1,20 @@ +--- +# Name of the organizer +name: Lucas Lehnert + +# Link to the organizer's webpage +webpage: https://lucaslehnert.github.io/ + +# Primary affiliation +affil: 2019 +# Link to the primary affiliation +# affil_link: https://buzz.edu + +# An image of the organizer (square aspect ratio works the best) (place in the `assets/img/organizers` directory) +img: lucaslehnert.jpeg + +# # Secondary affiliation +# affil2: BuzzFizz Corp +# # Link to the secondary affiliation +# affil2_link: https://buzzfizz.corp +--- diff --git a/_organizers/organizerjohndoe.md b/_organizers/organizerjohndoe.md deleted file mode 100644 index 0e2f3c0..0000000 --- a/_organizers/organizerjohndoe.md +++ /dev/null @@ -1,4 +0,0 @@ ---- -name: Organizer John Doe -affil: Fizz University ---- diff --git a/_speakers/1_stefanoalbrecht.md b/_speakers/1_stefanoalbrecht.md index 46204ac..9e8be1b 100644 --- a/_speakers/1_stefanoalbrecht.md +++ b/_speakers/1_stefanoalbrecht.md @@ -1,6 +1,6 @@ --- # Name of the speaker -name: Stefano Albrecht +name: Stefano V. Albrecht # Link to the speaker's webpage webpage: https://agents.inf.ed.ac.uk/stefano-albrecht/ diff --git a/assets/abstracts/timrudner.txt b/assets/abstracts/timrudner.txt deleted file mode 100644 index e69de29..0000000 diff --git a/assets/img/organizers/benjaminspiegel.png b/assets/img/organizers/benjaminspiegel.png new file mode 100644 index 0000000..b0f5989 Binary files /dev/null and b/assets/img/organizers/benjaminspiegel.png differ diff --git a/assets/img/organizers/kaiyuzheng.png b/assets/img/organizers/kaiyuzheng.png new file mode 100644 index 0000000..37f6a53 Binary files /dev/null and b/assets/img/organizers/kaiyuzheng.png differ diff --git a/assets/img/organizers/lucaslehnert.jpeg b/assets/img/organizers/lucaslehnert.jpeg new file mode 100644 index 0000000..4c9d9c1 Binary files /dev/null and b/assets/img/organizers/lucaslehnert.jpeg differ diff --git a/assets/img/organizers/sakettiwari.jpeg b/assets/img/organizers/sakettiwari.jpeg new file mode 100644 index 0000000..d174bda Binary files /dev/null and b/assets/img/organizers/sakettiwari.jpeg differ diff --git a/assets/img/organizers/samlobel.jpeg b/assets/img/organizers/samlobel.jpeg new file mode 100644 index 0000000..cfd7e20 Binary files /dev/null and b/assets/img/organizers/samlobel.jpeg differ diff --git a/assets/img/organizers/ziyiyang.jpeg b/assets/img/organizers/ziyiyang.jpeg new file mode 100644 index 0000000..9d03c3b Binary files /dev/null and b/assets/img/organizers/ziyiyang.jpeg differ diff --git a/contactus.md b/contactus.md index a9d7afb..69edced 100644 --- a/contactus.md +++ b/contactus.md @@ -4,4 +4,4 @@ title: Contact Us permalink: /contactus/ --- -Here's how you can contact us ... +Please feel free to email [us](mailto:ziyi_yang1@brown.edu) if you have any question or suggestion! diff --git a/index.md b/index.md index f7a5b66..2b1020a 100644 --- a/index.md +++ b/index.md @@ -23,12 +23,12 @@ Brown Robotics Talks consists of BigAI talks and lab talks (CIT 115). From sprin 02/02 - Deep Reinforcement Learning for Multi-Agent Interaction - Stefano Albrecht + Deep Reinforcement Learning for Multi-Agent Interaction + Stefano V. Albrecht 03/01 - TBD + TBD Felix Yanwei Wang diff --git a/pasttalks.md b/pasttalks.md index d312344..c0c9eb0 100644 --- a/pasttalks.md +++ b/pasttalks.md @@ -4,8 +4,8 @@ title: Past Talks permalink: /pasttalks/ --- -You'll need a Brown account to access the recordings. - +(You'll need a Brown account to access the recordings.) +

2023

@@ -15,8 +15,80 @@ You'll need a Brown account to access the recordings. - - + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Regularization in Neural Networks: A Probabilistic Perspective [recording]Tim RudnerRegularization in Neural Networks: A Probabilistic Perspective [abstract][recording]Tim Rudner [bio]
Broadening Robot Dexterity: Leveraging Elements in Manipulation Task Environments [abstract][recording]Xianyi Cheng [bio]
The Hessian Perspective into the Nature of Neural Networks [abstract][recording]Sidak Pal Singh [bio]
Quasimetric Reinforcement Learning [abstract][recording]Tongzhou Wang [bio]
Generalizing Beyond Your Data through Compositional Energy Functions [abstract][recording]Yilun Du [bio]
A2Perf: Benchmarking Autonomous Agents for the Real World [abstract][recording]Ike Uchendu [bio]
Towards Social Autotelic Agents: Open-Ended Skill Learning with Goals, Language and Intrinsically Motivated Reinforcement Learning [abstract][recording]Cédric Colas [bio]
Natural Language Task Specification for Robots [abstract][recording]Jason Liu [bio]
Deep Symbol Generation and Rule Learning [abstract][recording]Alper Ahmetoglu [bio]
Pragmatic, Uncertainty Guided Reinforcement Learning [abstract][recording]Taylor Killian [bio]
Grounded Understanding of Actions and Language via Bayesian Inverse Planning [abstract][recording]Tan Zhi-Xuan
Improving Unsupervised Visual Program Inference with Code Rewriting Families [abstract][recording]Aditya Ganeshan [bio]
Inventing Plannable Abstractions from Demonstrations [abstract][recording]Nishanth Kumar [bio]
Equivariant Learning for Robotic Manipulation [abstract][recording]Dian Wang [bio]
Scaling Goal-based Exploration via Pruning Proto-goals [abstract] & Learning with Program Induction [abstract][recording]Akhil Bagaria & Skye Thompson
Open and Efficient Reinforcement Learning from Human Feedback [abstract][recording]Louis Catricato [bio]
The Uninteded Consequences of Discount Regularization [recording]Sarah Rathnam [bio]
Interpretable Artificial Intelligence for Personalized Human-Robot Collaboration [recording]Rohan Paleja [bio]
Choreorobotics: An Emerging Discipline [abstract][recording]Catie Cuan [bio]
diff --git a/pasttalks/abstracts/adityaganeshan.txt b/pasttalks/abstracts/adityaganeshan.txt new file mode 100644 index 0000000..2e873b1 --- /dev/null +++ b/pasttalks/abstracts/adityaganeshan.txt @@ -0,0 +1 @@ +3D Data is often modeled with discrete symbolic structure to improve editability and interpretability. However, inferring this symbolic representation directly from the data is challenging - learning methods often suffer due to poor gradient estimation over stochastic computational graphs (especially with categorical distributions). An alternate strategy is to bootstrap a neurally guided search process by increasing the likelihood of “good” structures discovered during the search itself. In this talk, I will first present this approach, which we term “bootstrapped learning” and then present my recent work, which improves neurally guided inference of visual programs by code rewriting. Our work improves neurally guided visual program inference across different visual languages and matches the reconstruction accuracy of domain specific architectures while inferring concise and interpretable programs. Overall, the talk with present the challenges associated with inferring symbolic representations from 3D data, and how bootstrapped learning and code rewriting can aid with this task. \ No newline at end of file diff --git a/pasttalks/abstracts/akhilbagaria.txt b/pasttalks/abstracts/akhilbagaria.txt new file mode 100644 index 0000000..af2b3a9 --- /dev/null +++ b/pasttalks/abstracts/akhilbagaria.txt @@ -0,0 +1 @@ +One of the gnarliest challenges in reinforcement learning is exploration that scales to vast domains, where novelty-, or coverage-seeking behavior falls short. Goal-directed, purposeful behaviors are able to overcome this, but rely on a good goal space. The core challenge in goal discovery is finding the right balance between generality (not hand-crafted) and tractability (useful, not too many). Our approach explicitly seeks the middle ground, enabling the human designer to specify a vast but meaningful proto-goal space, and an autonomous discovery process to narrow this to a narrow space of controllable, reachable, novel, and relevant goals. The effectiveness of goal-conditioned exploration with the latter is then demonstrated in three challenging environments. \ No newline at end of file diff --git a/pasttalks/abstracts/alperahmetoglu.txt b/pasttalks/abstracts/alperahmetoglu.txt new file mode 100644 index 0000000..95baccb --- /dev/null +++ b/pasttalks/abstracts/alperahmetoglu.txt @@ -0,0 +1 @@ +Learning the symbolic representation of tasks enables the application of classical AI search techniques to find a solution in the symbolic definition of the task. For unknown, unstructured, and/or changing environments, it is desirable that the robot itself discovers the symbolic structures that are useful in reasoning and planning. In this talk, I will introduce our recent work, DeepSym, which aims to learn symbols with deep networks from unsupervised robot interactions and build rules defined over these symbols for domain-independent planning. I will continue with follow-up works that extend DeepSym to learn multi-object and relational symbols. I will conclude with a discussion on why an open-ended learning approach would be practical in symbol learning for truly intelligent robots. \ No newline at end of file diff --git a/pasttalks/abstracts/catiecuan.txt b/pasttalks/abstracts/catiecuan.txt new file mode 100644 index 0000000..960bf22 --- /dev/null +++ b/pasttalks/abstracts/catiecuan.txt @@ -0,0 +1 @@ +Choreographers and roboticists both organize moving bodies in space, yet the platforms and practices they employ are seemingly dissimilar. This talk describes the emerging field of choreorobotics, from dancing robots to novel programming interfaces, and why the intersection of these artistic and engineering fields is rapidly expanding. \ No newline at end of file diff --git a/pasttalks/abstracts/cedriccolas.txt b/pasttalks/abstracts/cedriccolas.txt new file mode 100644 index 0000000..75241f7 --- /dev/null +++ b/pasttalks/abstracts/cedriccolas.txt @@ -0,0 +1 @@ +In this talk I will take a developmental perspective on the problem of modeling open-ended skill learning in artificial agents. I’ll argue that such agents need to be both autotelic and social — i.e. intrinsically motivated to represent and pursue their own goals, but still learning within human cultures. I will develop this argument by presenting several learning architecture I developed during my PhD. In the second part of the talk, I’ll discuss more recent and ongoing projects that aim at leveraging natural human feedback to teach artificial agents in an efficient way by leveraging a program-induction perspective on learning and reasoning. \ No newline at end of file diff --git a/pasttalks/abstracts/dianwang.txt b/pasttalks/abstracts/dianwang.txt new file mode 100644 index 0000000..ecae694 --- /dev/null +++ b/pasttalks/abstracts/dianwang.txt @@ -0,0 +1 @@ +In the field of robotics, researchers have frequently exploited geometric and physical structures to simplify complex planning and control problems. One such natural geometric structure is symmetry. Recently, equivariant models, which encode symmetries within the neural network architecture, have achieved remarkable success in the machine learning community. This raises the question of how these geometric deep learning techniques can be effectively applied to robotics. Our recent work has taken a step in this direction by applying equivariant learning methods to robotic manipulation, aiming to enhance sample efficiency and generalization. Our findings reveal that these methods significantly outperform conventional unconstrained models and effectively enable on-robot reinforcement learning. diff --git a/pasttalks/abstracts/ikeuchendu.txt b/pasttalks/abstracts/ikeuchendu.txt new file mode 100644 index 0000000..4bb9aff --- /dev/null +++ b/pasttalks/abstracts/ikeuchendu.txt @@ -0,0 +1 @@ +Benchmarks have proven integral for driving progress in artificial intelligence across areas like computer vision and natural language processing. However, evaluating autonomous agents for real-world applications poses challenges including system constraints, generalization, and reliability. In this talk, I introduce A2Perf, a benchmarking suite currently under development to evaluate agent performance on tasks derived from real-world domains. The design of A2Perf specifically targets metrics that reflect challenges observed in practice, such as inference latency, memory usage, and generalizability. I provide an overview of A2Perf's proposed domains including computer chip-floorplanning, web navigation, and quadruped locomotion. I will discuss the current status of this ongoing effort, our design goals, and future directions. We believe tailoring benchmark tasks and metrics to real-world needs in this way will ultimately help guide and accelerate research on deployable autonomous agents. \ No newline at end of file diff --git a/pasttalks/abstracts/jasonliu.txt b/pasttalks/abstracts/jasonliu.txt new file mode 100644 index 0000000..6ae14dd --- /dev/null +++ b/pasttalks/abstracts/jasonliu.txt @@ -0,0 +1 @@ +Natural language provides an intuitive and flexible way for human users to specify tasks to robots. Linear temporal logic (LTL) provides unambiguous semantics for natural language grounding, and its compositional grammar can induce skill transfer. However, existing methods are limited in grounding natural language commands in unseen environments, and existing learning algorithms for temporal tasks only support limited policy reuse for novel specifications. This thesis proposes an integrated robotic system from natural language task specification to robot actions that is generalizable to unseen environments and tasks. We first introduce a modular system, Lang2LTL, that uses pre-trained large language models to ground temporally-extended natural language commands to LTL expressions. Lang2LTL achieves higher accuracy than previous state-of-the-art methods in grounding complex temporal task specifications in unseen environments of city and house scales. We then propose LTL-Transfer, an algorithm that leverages the compositional nature of LTL to reuse learned policies to solve novel tasks in zero-shot. We deployed our integrated system on a physical robot to solve temporal tasks specified by natural language in zero-shot. Our ongoing work is investigating using human-robot dialog to resolve ambiguity in grounding natural language task specification. \ No newline at end of file diff --git a/pasttalks/abstracts/louiscastricato.txt b/pasttalks/abstracts/louiscastricato.txt new file mode 100644 index 0000000..f453013 --- /dev/null +++ b/pasttalks/abstracts/louiscastricato.txt @@ -0,0 +1 @@ +Over the past couple months CarperAI has built trlX, one of the first open source RLHF implementations capable of fine-tuning large language models at scale. We test offline reinforcement algorithms to reduce compute requirements and explore the practicality of synthetic preference data, finding both can be combined to significantly reduce expensive RLHF costs. \ No newline at end of file diff --git a/pasttalks/abstracts/nishanthkumar.txt b/pasttalks/abstracts/nishanthkumar.txt new file mode 100644 index 0000000..7cf2fe7 --- /dev/null +++ b/pasttalks/abstracts/nishanthkumar.txt @@ -0,0 +1 @@ +Although ML and AI have advanced considerably, robots still struggle with performing practical household tasks, such as making tea in any kitchen. Recent work has shown that combining learning and planning by learning structured world models from data is a promising path towards enabling this capability. In this talk, I'll discuss our recent efforts to extend learning for planning approaches to the task and motion planning (TAMP) setting. I will present methods that only require a handful of demonstrations to invent abstractions in the form of symbolic predicates and operators, as well as neural samplers, to enable TAMP. I will show how these learned components not only support efficient decision-making via planning, but also zero-shot generalization to more challenging tasks. Finally, I will highlight the current methods' limitations and identify critical areas for future work necessary to scale this approach to physical robots tackling real-world tasks. \ No newline at end of file diff --git a/pasttalks/abstracts/sidakpalsingh.txt b/pasttalks/abstracts/sidakpalsingh.txt new file mode 100644 index 0000000..6300875 --- /dev/null +++ b/pasttalks/abstracts/sidakpalsingh.txt @@ -0,0 +1 @@ +The Hessian of a neural network captures parameter interactions through second-order derivatives of the loss. It is a fundamental object of study, closely tied to various problems in deep learning, including model design, optimization, and generalization. Most prior work has been empirical, typically focusing on low-rank approximations and heuristics that are mostly blind to the network structure. In contrast, we develop theoretical tools to analyze the range of the Hessian map, which provides us with a precise understanding of its rank deficiency and the structural reasons behind it. This yields exact formulas and tight upper bounds for the Hessian rank of deep linear networks --- allowing for an elegant interpretation in terms of rank deficiency. Further, by relying on a Toeplitz representation, we extend our results to the case of deep linear convolutional networks. Moreover, we demonstrate that our bounds remain faithful as an estimate of the numerical Hessian rank, for a larger class of networks with non-linearities such as rectified and hyperbolic tangent networks. Overall, our work generalizes and further establishes the key insight that the Hessian rank grows as the square root of the number of parameters --- thus providing novel insights into the source and extent of redundancy in overparameterized neural networks. \ No newline at end of file diff --git a/pasttalks/abstracts/skyethompson.txt b/pasttalks/abstracts/skyethompson.txt new file mode 100644 index 0000000..8e6cbdf --- /dev/null +++ b/pasttalks/abstracts/skyethompson.txt @@ -0,0 +1 @@ +What if our agents (or robots, or classifiers, etc) could just program themselves? Program induction is a set of techniques for learning programs from data - building higher-level function representations from a lower-level language that can encode useful priors or constraints in a specific domain. In recent years, they’ve been applied in RL and robotics for learning policies, operators, transition conditions, object representations and more. How do they work, why would you use them, and maybe most importantly, when should you skip them entirely? This is a discussion focused presentation including an intro to program induction techniques, a review of some recent work in RL and robotics using them, and some of the presenter’s work to give a picture of the strengths and limitations of this family of learning approaches. \ No newline at end of file diff --git a/pasttalks/abstracts/tanzhixuan.txt b/pasttalks/abstracts/tanzhixuan.txt new file mode 100644 index 0000000..d87a42d --- /dev/null +++ b/pasttalks/abstracts/tanzhixuan.txt @@ -0,0 +1,15 @@ +People routinely infer the goals of others from both actions and words, understanding not just their surface content, but also the intentions and desires that underlie them. How might we build assistive machines that do the same? This talk will first introduce Bayesian inverse planning as a general framework for goal inference problems. I will then show how these problems can be solved accurately and efficiently via sequential inverse plan/policy search (SIPS), a family of algorithms that models agents as online model-based planners, and uses programmable sequential Monte Carlo to rapidly infer agents' goals and plans from observations of their behavior. Both planning and inference are grounded in PDDL representations of the environment via PDDL.jl, a extensible compiler for classical and numeric PDDL planning with comparable performance to FastDownward and ENHSP. This delivers speed, generality, and interpretability, while also enabling assistive robots to plan helpful actions. + +Because SIPS is highly configurable, it can be used to model boundedly-rational agents: Agents that plan ahead only a few steps at a time, and hence might exhibit mistakes from not planning enough. This allows us to infer an agent's goals even when they fail to achieve them, accurately mirroring how humans infer others' goals from their mistaken plans, while significantly outperforming Boltzmann-rational agent models. SIPS can also be extended to handle linguistic input: By modeling how an agent not only acts, but also communicates rationally to achieve their goals, we can infer the pragmatic intentions behind their utterances. We do this by integrating (large) language models (LLMs) as likelihood functions over utterances, allowing us to handle a wide range of utterance forms, and to infer an agent's goal from both their actions and utterances. + +Finally, I will present ongoing work on how Bayesian inverse planning can be extended to even richer settings. In environments where human goals exhibit rich compositional structures (e.g. cooking a meal in a kitchen), I will show how LLMs can be used as common sense goal priors, enabling accurate goal inference even when there are millions of semantically valid goals. In continuous environments, I will also show how goal inference can be performed via Bayesian inverse motion planning, avoiding the limits of discretization. These advances pave the way towards fast, flexible, and grounded inferences over the infinite variety of human goals, furthering the development of human-aligned assistive systems. + +Relevant papers: +Online Bayesian Goal Inference for Boundedly Rational Planning Agents +Modeling the Mistakes of Boundedly Rational Agents with an Bayesian Theory of Mind +Inferring the Goals of Communicating Agents from Actions and Instructions + +Relevant libraries: +https://github.com/probcomp/GenParticleFilters.jl +https://github.com/JuliaPlanners/PDDL.jl +https://github.com/JuliaPlanners/SymbolicPlanners.jl diff --git a/pasttalks/abstracts/taylorkillian.txt b/pasttalks/abstracts/taylorkillian.txt new file mode 100644 index 0000000..1e52376 --- /dev/null +++ b/pasttalks/abstracts/taylorkillian.txt @@ -0,0 +1 @@ +The development of a successful RL policy rests on the ability to derive informative states from observations and explore alternative strategies from time to time. In many real-world scenarios, such as healthcare, these observations are noisy, irregular, and may not convey all salient information to form a decision. Additionally current state-of-the-art RL algorithms, when faced with partial information and the inability to proactively experiment or explore within their environment, fail to reliably learn optimal policies. With limited data in such settings, determining an optimal policy is intractable. However, recorded negative outcomes can still be useful to identify behaviors that should be avoided. In this talk, I will highlight specific modeling decisions that can be made to develop actionable insights from sequentially observed healthcare data to facilitate the avoidance of suboptimal decisions in patient care. These modeling choices honor underlying data generation as well as the processes by which clinical experts use to formulate their own decisions. \ No newline at end of file diff --git a/pasttalks/abstracts/timrudner.txt b/pasttalks/abstracts/timrudner.txt new file mode 100644 index 0000000..2dc3ebe --- /dev/null +++ b/pasttalks/abstracts/timrudner.txt @@ -0,0 +1,7 @@ +Conventional regularization techniques for neural networks, such as L2 or L1 regularization, explicitly penalize divergence of the model parameters from specific parameter values. However, in most neural network models, specific parameter configurations bear little to no physical meaning, and it is difficult to incorporate domain knowledge or other relevant information into neural network training using conventional regularization techniques. + +In this talk, I will show that we can address this shortcoming by using Bayesian principles to effectively incorporate domain knowledge or beliefs about desirable model properties into neural network training. To do so, I will approach regularization in neural networks from a probabilistic perspective and define a family of data-driven prior distributions that allows us to encode useful auxiliary information into the model. I will then show how to perform approximate inference in neural networks with such priors and derive a simple variational optimization objective with a regularizer that reflects the constraints implicitly encoded in the prior. This regularizer is mathematically simple, easy to implement, and can be used as a drop-in replacement for existing regularizers when performing supervised learning in neural networks of any size. + +I will conclude the talk with an overview of applications of data-driven priors, including distribution shift detection and medical diagnosis. + +This is joint work with Sanyam Kapoor, Shikai Qiu, Xiang Pan, Lily Yucen Li, Ya Shi Zhang, Ravid Shwartz-Ziv, Julia Kempe, and Andrew Gordon Wilson. \ No newline at end of file diff --git a/pasttalks/abstracts/tongzhouwang.txt b/pasttalks/abstracts/tongzhouwang.txt new file mode 100644 index 0000000..d4e498f --- /dev/null +++ b/pasttalks/abstracts/tongzhouwang.txt @@ -0,0 +1,4 @@ +In goal-reaching agents, how are strategies for different goals related? Can we solve goal-reaching reinforcement learning (RL) with a sufficiently good representation of states and goals? In this talk, I will present a method for training high-performance optimal goal-reaching agents by learning a quasimetric geometry. This talk consists of three parts: +1. Goal-Reaching RL == Quasimetric geometry learning. +2. How to represent this geometry? Deep quasimetric models. +3. How to learn this geometry from local transitions? A geometric argument based on quasimetric properties. \ No newline at end of file diff --git a/pasttalks/abstracts/xianyicheng.txt b/pasttalks/abstracts/xianyicheng.txt new file mode 100644 index 0000000..aa5ebe4 --- /dev/null +++ b/pasttalks/abstracts/xianyicheng.txt @@ -0,0 +1,3 @@ +I believe robot dexterity is not only about the robot's physical flexibility (characterized by more joints), but instead, we should consider it as a critical form of artificial intelligence. This talk presents some of my thoughts about opportunities we could open up about robot dexterity. + +My past work discusses a question, " Where could a robot discover more dexterity?". I will present the idea that we could expand the frontier of robot dexterity by using all elements present in the task environments: not only the robot fingertips but also the environment and other existing features of the robot. For example, a robot can find dexterity in even the simplest gripper, a suction cup designed solely for pick-and-place. By fully considering the suction cup soft bellow and exploiting the environment contacts, the robot can reorient an object, even in a way that a rigid robot finger cannot do. Second, the surrounding environment is an effective tool to enhance dexterity. I will present my series of works that generate various dexterous manipulation strategies by considering the present environment, without using any manually engineered primitives. Finally, I hope to discuss with you broadly about new opportunities for dexterity I discovered in this journey. \ No newline at end of file diff --git a/pasttalks/abstracts/yilundu.txt b/pasttalks/abstracts/yilundu.txt new file mode 100644 index 0000000..e87c0c0 --- /dev/null +++ b/pasttalks/abstracts/yilundu.txt @@ -0,0 +1,7 @@ +Generative AI has led to stunning successes in recent years but is fundamentally limited by the amount of data available. +This is especially challenging in robotics, where data is often missing and especially difficult to acquire. In this talk, I’ll introduce +the idea of compositional generative modeling, which can significantly reduce data requirements by building complex generative models +from smaller constituents. First, I introduce the idea of modeling generative models with energy landscapes and illustrate how they enable +compositional generative modeling. I’ll then illustrate how such compositional models enable the synthesis of robotic plans for novel environments + as well as complex visual scenes. Finally, I'll show how such compositionality can be applied to multiple “foundation models” to construct robotic +systems that can hierarchically plan and reason with multimodal inputs to solve long-horizon problems. \ No newline at end of file diff --git a/pasttalks/bios/adityaganeshan.txt b/pasttalks/bios/adityaganeshan.txt new file mode 100644 index 0000000..73b89ca --- /dev/null +++ b/pasttalks/bios/adityaganeshan.txt @@ -0,0 +1 @@ +Aditya Ganeshan is a second-year Ph.D. student at the Brown visual computing group, Brown University advised by Professor Daniel Ritchie. He is interested in developing techniques for data-driven inference of structured representations for data (typically 3D data). Previously, he worked on computer vision research at Preferred Networks Inc., Tokyo, Japan, and adversarial robustness of neural networks at IISc, Banglore, India . He completed his Bachelor’s & Master’s degree in Applied Mathematics from IIT Roorkee, India. \ No newline at end of file diff --git a/pasttalks/bios/alperahmetoglu.txt b/pasttalks/bios/alperahmetoglu.txt new file mode 100644 index 0000000..ee0f0fb --- /dev/null +++ b/pasttalks/bios/alperahmetoglu.txt @@ -0,0 +1 @@ +Alper Ahmetoglu is a PhD candidate at Bogazici University supervised by Emre Ugur and Erhan Oztop. He received his Bachelor's and Master's degrees from Bogazici University in 2017 and 2019, respectively. In his master's studies, he worked with Ethem Alpaydin on hierarchical mixtures of experts and generative adversarial nets. His research interests are open-ended learning, neurosymbolic methods, and affordances for the ultimate purpose of creating intelligent agents. \ No newline at end of file diff --git a/pasttalks/bios/catiecuan.txt b/pasttalks/bios/catiecuan.txt new file mode 100644 index 0000000..a18a446 --- /dev/null +++ b/pasttalks/bios/catiecuan.txt @@ -0,0 +1 @@ +Catie Cuan is a PhD Candidate in the Mechanical Engineering department at Stanford University, where she completed a Master's of Science in Mechanical Engineering. Her research focuses on robotics, haptics, and imitation learning and she has received fellowships from the National Institutes of Health, Everyday Robots, and Stanford. A professional dancer and choreographer, her artistic work is situated between dance and robotics. She has been awarded artist residencies at TED, ThoughtWorks Arts, and the Smithsonian. She is currently the inaugural Artist-in-Residence at Everyday Robots, from Google X, the moonshot factory. diff --git a/pasttalks/bios/cedriccolas.txt b/pasttalks/bios/cedriccolas.txt new file mode 100644 index 0000000..529ff61 --- /dev/null +++ b/pasttalks/bios/cedriccolas.txt @@ -0,0 +1 @@ +Colas is a postdoctoral researcher in the Computational Cognitive Science Lab at MIT and the Flowers Lab at INRIA. His PhD thesis introduced the concept of autotelic learning: a computational framework for endowing artificial agents with the ability to represent and pursue their own goals. Through the co-evolution of goal generation and goal achievement systems, autotelic learning makes the first steps towards a computational model of human’s ability to grow repertoires of skills throughout a lifetime. His current research interests include the examination of the role of social interactions in shaping the goal representations and goal reaching skills of artificial agents. \ No newline at end of file diff --git a/pasttalks/bios/dianwang.txt b/pasttalks/bios/dianwang.txt new file mode 100644 index 0000000..e5de2b1 --- /dev/null +++ b/pasttalks/bios/dianwang.txt @@ -0,0 +1 @@ +Dian Wang is a fourth-year Ph.D. candidate at the Khoury College of Computer Science at Northeastern University, advised by Professor Robert Platt and Professor Robin Walters. His research interests include Robotic Manipulation, Equivariant Machine Learning, and Reinforcement Learning. His research focuses on applying equivariant learning methods in robotic manipulation to improve sample efficiency and generalization. Prior to his Ph.D., he received a Master's degree in Computer Science from Northeastern University and a Bachelor's degree in Computer Science and Engineering from Sichuan University. \ No newline at end of file diff --git a/pasttalks/bios/ikeuchendu.txt b/pasttalks/bios/ikeuchendu.txt new file mode 100644 index 0000000..ece67e0 --- /dev/null +++ b/pasttalks/bios/ikeuchendu.txt @@ -0,0 +1 @@ +Ikechukwu "Ike" Uchendu is a PhD student at Harvard's Edge Computing Lab and a Student Researcher at Google DeepMind. His research is centered on the development and evaluation of autonomous agents for real-world applications. Before this, he was an AI Resident at Google Brain, focusing on reinforcement learning for robotic manipulation. Ike earned his BS and MS degrees in Computer Science & Engineering from Michigan State University in 2018 and 2020. He has published papers in conferences such as ICML and ISCA, as well as workshops at NeurIPS and ICRA. \ No newline at end of file diff --git a/pasttalks/bios/jasonliu.txt b/pasttalks/bios/jasonliu.txt new file mode 100644 index 0000000..b9ee44f --- /dev/null +++ b/pasttalks/bios/jasonliu.txt @@ -0,0 +1 @@ +Jason Liu is a Ph.D. candidate at Brown University, advised by Prof. Stefanie Tellex. His work lies in the intersection of robotics and natural language processing. He is building robotic systems that understand natural language commands specifying long-horizon temporal tasks. Previously, Jason earned his Bachelor's degree in Electrical Engineering and Computer Sciences from UC Berkeley. He is fortunately being funded by NSF Graduate Research Fellowship Program and Jack Kent Cooke Foundation Graduate Scholarship. \ No newline at end of file diff --git a/pasttalks/bios/louiscastricato.txt b/pasttalks/bios/louiscastricato.txt new file mode 100644 index 0000000..743451c --- /dev/null +++ b/pasttalks/bios/louiscastricato.txt @@ -0,0 +1 @@ +Louis is a PhD student at Brown, studying under Professor Ellie Pavlick. Louis also works at Stability AI as a research scientist and is team lead at CarperAI. Previously, he was a research scientist at EleutherAI. His background is in preference modeling and computational narrative theory. \ No newline at end of file diff --git a/pasttalks/bios/nishanthkumar.txt b/pasttalks/bios/nishanthkumar.txt new file mode 100644 index 0000000..dd286a5 --- /dev/null +++ b/pasttalks/bios/nishanthkumar.txt @@ -0,0 +1 @@ +Nishanth is a 2nd year Ph.D. student in the LIS Group at MIT CSAIL, where his research is supported by an NSF GRFP fellowship. He is primarily interested in enabling robots to operate robustly in long-horizon, multi-task settings so that they can accomplish tasks like multi-object manipulation, cooking, or even performing household chores. To this end, his work seeks to synthesize ideas from a number of sub-fields of AI, including Task and Motion Planning, Reinforcement Learning, Program Synthesis and Neurosymbolic methods. Previously, Nishanth obtained a Bachelor of Science in Computer Engineering from Brown University, where he was a Goldwater Scholar, CRA Outstanding Undergrad Researcher Award Finalist, and was named the Outstanding Senior in Computer Engineering upon graduation. \ No newline at end of file diff --git a/pasttalks/bios/rohanpaleja.txt b/pasttalks/bios/rohanpaleja.txt new file mode 100644 index 0000000..e6171b3 --- /dev/null +++ b/pasttalks/bios/rohanpaleja.txt @@ -0,0 +1 @@ +Rohan R Palej is a Robotics Ph.D. candidate at Georgia Tech. His research covers a diverse set of topics, including multi-agent systems, human-robot teaming, explainable AI, and Interactive Robot Learning. His goal is to create intelligent, transparent robotic teammates that can anticipate human behavior, communicate with human teammates to increase user awareness, and collaborate with high performance. \ No newline at end of file diff --git a/pasttalks/bios/sarahrathnam.txt b/pasttalks/bios/sarahrathnam.txt new file mode 100644 index 0000000..479ec28 --- /dev/null +++ b/pasttalks/bios/sarahrathnam.txt @@ -0,0 +1 @@ +Sarah is a PhD student in Applied Mathematics at Harvard advised by Finale Doshi-Velez of the Data to Actionable Knowledge Lab and Susan Murphy of the Statistical Reinforcement Learning Lab. She focuses on reinforcement learning for mobile health, specifically on issues regarding regularization and balancing short term and long term goals. In her previous life, she worked as a quant trader at various hedge funds and proprietary trading funds. \ No newline at end of file diff --git a/pasttalks/bios/sidakpalsingh.txt b/pasttalks/bios/sidakpalsingh.txt new file mode 100644 index 0000000..e65bf29 --- /dev/null +++ b/pasttalks/bios/sidakpalsingh.txt @@ -0,0 +1 @@ +Sidak Pal Singh is a 4th year PhD student at ETH Zürich and Max Planck Insitute of Intelligent Systems, Tübingen advised by Thomas Hofmann and Bernhard Schölkopf. Before that, he obtained his Masters in Data Science at EPFL and Bachelor in Computer Science at the Indian Insitute of Technology Roorkee. Sidak's main research interests are in uncovering structural and functional properties of neural networks, with the larger aim of developing a better and more precise understanding of the remarkable generalization capabilities of neural networks. This includes, among other things, questions surrounding the loss landscape and its curvature, double descent behaviour of generalization, mode connectivity, functional smoothness and robustness (Lipschitz continuity). \ No newline at end of file diff --git a/pasttalks/bios/taylorkillian.txt b/pasttalks/bios/taylorkillian.txt new file mode 100644 index 0000000..58d54b6 --- /dev/null +++ b/pasttalks/bios/taylorkillian.txt @@ -0,0 +1 @@ +Taylor is in his final year of a PhD program in Computer Science at University of Toronto, with standing affiliations at Vector Institute and MIT Institute of Medical Engineering and Sciences. His research broadly investigates novel applications of Reinforcement Learning to assist sequential decision making in safety-critical domains. In particular, he is interested in developing personalized decision support tools that generalize beyond the environment they were trained in, robust to sources of uncertainty such as distribution shift, covariate mismatch, and missing data. Taylor has prior degrees in Computational Science and Engineering (M.Eng, Harvard University) as well as Mathematics (B.S. Brigham Young University) that he has used in prior research positions investigating the RF scheduling algorithms, sensor placement (MIT Lincoln Laboratory), and in the modeling of fluid phenomenon (BYU R.A.). \ No newline at end of file diff --git a/pasttalks/bios/timrudner.txt b/pasttalks/bios/timrudner.txt new file mode 100644 index 0000000..a706e10 --- /dev/null +++ b/pasttalks/bios/timrudner.txt @@ -0,0 +1 @@ +Tim G. J. Rudner is an Assistant Professor and Faculty Fellow at New York University’s Center for Data Science and an AI Fellow at Georgetown University's Center for Security and Emerging Technology. He conducted PhD research on probabilistic machine learning in the Department of Computer Science at the University of Oxford, where he was advised by Yee Whye Teh and Yarin Gal. The goal of his research is to create trustworthy machine learning models by developing methods and theoretical insights that improve the reliability, safety, transparency, and fairness of machine learning systems deployed in safety-critical settings. Tim holds a master’s degree in statistics from the University of Oxford and an undergraduate degree in applied mathematics and economics from Yale University. He is also a Rhodes Scholar and a Qualcomm Innovation Fellow. \ No newline at end of file diff --git a/pasttalks/bios/tongzhouwang.txt b/pasttalks/bios/tongzhouwang.txt new file mode 100644 index 0000000..8267155 --- /dev/null +++ b/pasttalks/bios/tongzhouwang.txt @@ -0,0 +1 @@ +Tongzhou is a final year PhD student at MIT, advised by Phillip Isola and Antonio Torralba. His research interests lie in structures in machine learning and artificial agents, focusing on learning structured representations for better perception and decision-making. His work spans representation learning, reinforcement learning, and machine learning. Tongzhou co-organized the Goal-Conditioned Reinforcement Learning workshop at NeurIPS 2023, bridging researchers and practitioners across machine learning and decision-making. Before his PhD study, Tongzhou received his bachelor's degree from UC Berkeley while working with Stuart Russell, Alyosha Efros and Ren Ng, and was an early member of the PyTorch team at Facebook AI Research. \ No newline at end of file diff --git a/pasttalks/bios/xianyicheng.txt b/pasttalks/bios/xianyicheng.txt new file mode 100644 index 0000000..72e3b3f --- /dev/null +++ b/pasttalks/bios/xianyicheng.txt @@ -0,0 +1 @@ +Xianyi Cheng is a Ph.D. candidate at Carnegie Mellon University, advised by Professor Matt Mason. Her primary research interests are in robotic manipulation and manipulation dexterity. Specifically, her current work focuses on the automatic generation and planning of versatile dexterous manipulation skills. She has received the Foxconn Graduate Fellowship and was selected to participate in the MIT EECS Rising Star in 2021. \ No newline at end of file diff --git a/pasttalks/bios/yilundu.txt b/pasttalks/bios/yilundu.txt new file mode 100644 index 0000000..b30bf67 --- /dev/null +++ b/pasttalks/bios/yilundu.txt @@ -0,0 +1,3 @@ +Yilun Du is PhD student at MIT EECS advised by Leslie Kaelbling, Tomas Lozano-Perez, and Joshua Tenenbaum. Previously, +he has done a research fellowship at OpenAI, and held visiting researcher/research internship positions at FAIR and Google Deepmind. + He is supported by the NSF Graduate Fellowship, and has received outstanding paper awards at NeurIPS and ICRA workshops. \ No newline at end of file