Forum for Artificial Intelligence

Yoav Wald [homepage]
Faculty Fellow, New York University

Title: Robustness to Spurious Correlations: Lessons Learned on Causal Thinking and Large Scale Modelling

Abstract: Predictive models often rely on spurious correlations, leading to unreliable performance. This talk covers a series of works on causally motivated methods to train models that are robust to spurious correlations. I will begin by outlining a formal framework for reasoning about spurious correlations and cover some core techniques to mitigate their effects. The talk will then explore three aspects of developing robust models: (i) considerations in training overparameterized models that maintain invariance to shifts in spurious correlations, (ii) approaches to dealing with scenarios where annotations of spuriously correlated features are unavailable, and (iii) the use of large language models and causal estimation techniques to enhance robustness in large-scale, real-world medical text classification tasks.

About the speaker: Yoav Wald is a Faculty Fellow/Assistant Professor at NYU's Center for Data Science, working on causality-motivated machine learning and causal inference. He applies these tools to problems in robust prediction and decision support in healthcare. His interests include out-of-distribution generalization, novelty detection, sequential decision making, and interpretability. He received his PhD from the Hebrew University, was a long term research intern at Google Research and a postdoctoral fellow at Johns Hopkins University.

Robert Platt [homepage]
Associate Professor, Northeastern University

Title: Symmetric Policy Learning in Robotics

Abstract: Many robotics problems have transition dynamics that are symmetric in SE(2) and SE(3) with respect to rotation, translation, scaling, reflection, etc. In these situations, any optimal policy will also be symmetric over these transformations. In this talk, I leverage this insight to improve the data efficiency of policy learning by encoding domain symmetries directly into the neural network model using group invariant and equivariant layers. The result is that we can learn non-trivial visuomotor control policies with much less data than is typically the case. For imitation learning, this significantly reduces the number of demonstrations required. For reinforcement learning, it reduces the amount of experience needed to learn a good policy. In fact, we can sometimes learn good policies from scratch training directly on physical robotic hardware in real time.

About the speaker: Rob Platt is an Associate Professor in the Khoury College of Computer Sciences at Northeastern University and a Faculty Fellow at BDAII. He is interested in developing robots that can perform complex manipulation tasks alongside humans in the uncertain everyday world. Much of his work is at the intersection of robotic policy learning, planning, and perception. Prior to coming to Northeastern, he was a Research Scientist at MIT and a technical lead at NASA Johnson Space Center.
 

 
Daniel Fried [homepage]
Assistant Professor, Carnegie Mellon University

Title: Planning and Inferring With LLMs for Grounded, Interactive Tasks

Abstract: Large language models (LLMs) are increasingly being used in language-based interfaces for tasks that require interacting with people and with the (digital) world: for example asking questions of a user to help them interactively retrieve information, or carrying out everyday tasks in a web browser. These settings involve uncertainty and partial observability, and so afford planning and inference methods inspired by classical AI approaches. We present techniques for layering structured search and inference procedures on top of LLM-based agentive systems --- making the LLMs better able to interact with their environments and human partners. First, we investigate visually grounded reference games, where a system and a person must use dialogue to collaboratively identify and build common ground. Here, modeling uncertainty about the user's intent and asking maximally-informative questions improves task success. Second, we investigate language-based tasks on the web, where a system must carry out instructions from a person by taking actions in the browser. Here, performing tree search to plan over possible trajectories in the environments substantially improves performance of state-of-the-art models. 

About the speaker: Daniel Fried is an assistant professor in the Language Technologies Institute at CMU, and a research scientist at Meta AI. His research focuses on language grounding, interaction, and applied pragmatics, with a particular focus on language interfaces such as grounded instruction following and code generation. Previously, he was a postdoc at Meta AI and the University of Washington and completed a PhD at UC Berkeley. His research has been supported by an Okawa Research Award, a Google PhD Fellowship and a Churchill Fellowship

Bo Liu [homepage]
Ph.D. Student & Graduate Research Assistant, University of Texas at Austin

Title: Longhorn: State Space Models are Amortized Online Learners

Abstract: The most fundamental capability of modern AI methods such as Large Language Models (LLMs) is the ability to predict the next token in a long sequence of tokens, known as ``sequence modeling." Although the Transformers model is the current dominant approach to sequence modeling, its quadratic computational cost with respect to sequence length is a significant drawback. State-space models (SSMs) offer a promising alternative due to their linear decoding efficiency and high parallelizability during training. However, existing SSMs often rely on seemingly ad hoc linear recurrence designs. In this work, we explore SSM design through the lens of online learning, conceptualizing SSMs as meta-modules for specific online learning problems. This approach links SSM design to formulating precise online learning objectives, with state transition rules derived from optimizing these objectives. Based on this insight, we introduce a novel deep SSM architecture based on the implicit update for optimizing an online regression objective. Our experimental results show that our models outperform state-of-the-art SSMs, including the Mamba model, on standard sequence modeling benchmarks and language modeling tasks.

About the speaker: Bo is a PhD student at the University of Texas at Austin, advised by Prof. Peter Stone and Prof. Qiang Liu. His research lies in multitask/continual learning and reinforcement learning. Recently, he is working on designing theoretically sound algorithms/neural architectures for training large-scale multi-purpose agents.

Jason Liu [homepage]
Ph.D. Candidate, Brown University

Title: Robotic Language Grounding

Abstract: Natural language provides an intuitive and flexible way for humans to communicate with robots. However, understanding diverse, ambiguous language commands is challenging. Grounding language to structured task specifications enables autonomous robots to understand a broad range of natural language and solve long-horizon tasks with safety guarantees. Linear temporal logic (LTL) provides unambiguous semantics for language grounding, and its compositionality can induce skill transfer.

In this talk, I will first propose two language grounding systems. 1) Lang2LTL is a modular system that uses large language models (LLMs) to ground navigation commands with diverse temporal patterns to LTL task specifications in novel environments without retraining. 2) Improved upon its predecessor, Lang2LTL-2 uses LLMs and a pretrained vision-language model to ground spatiotemporal navigation commands. By translating language to LTL, both systems can detect infeasible task specifications and abort execution when necessary. 3) Next, I will introduce LTL-Transfer, a zero-shot transfer algorithm that leverages the compositionality of LTL to reuse learned skills to solve novel tasks without violating any safety constraints. Future work aims at developing robotic systems that produce robust and verifiable behavior by using multimodal grounding and human-robot dialog.

Relevant Papers:
https://arxiv.org/abs/2302.11649
https://spatiotemporal-ground.github.io/
https://arxiv.org/abs/2206.05096
https://arxiv.org/abs/2405.13245

About the speaker: Jason Xinyu Liu is a Ph.D. candidate at Brown University, advised by Prof. Stefanie Tellex. His research lies in the intersection of robotics, natural language processing, and formal methods. He is working towards developing autonomous robots that assist people. His work has appeared at CoRL, ICRA, IROS, IJCAI, and AAAI Symposiums. Jason earned his Bachelor's degree in Electrical Engineering and Computer Sciences from UC Berkeley. His research is generously funded by the NSF Graduate Research Fellowship Program and the Jack Kent Cooke Foundation Graduate Scholarship.

Leqi Liu [homepage]
Assistant Professor, The University of Texas at Austin

Title: Preference Optimization in Large Language Model Alignment: Personalization, Common Pitfalls and Beyond

Abstract: Reinforcement Learning from Human Feedback (RLHF) has become the predominant method for aligning large language models (LLMs) to be more helpful and less harmful. In this talk, we address two core limitations of traditional RLHF. First, it assumes that all human preferences come from the same distribution, preventing fine-tuned LLMs from generating personalized content without explicit prompting. We introduce Personalized RLHF, an efficient framework that captures individual preferences through a lightweight user model, enabling LLMs to generate content that reflects diverse and potentially conflicting user preferences. Second, current RLHF methods often rely on optimizing against margin-based losses, which focus on the difference between preferred and dispreferred responses but fail to specify ideal LLM behavior on each type of the responses individually. This underspecification can lead to problematic training dynamics, increasing the probability of generating unsafe content or reducing the probability of generating ideal responses. We characterize when these problematic dynamics emerge and outline algorithms that can mitigate these issues. Finally, we will discuss future directions and potential new paradigms for improving LLM alignment.

About the speaker: Leqi Liu is an assistant professor in use-inspired AI at the department of information, risk and operations management at UT Austin. Her research focuses on (1) investigating the foundations of state-of-the-art machine intelligence, with a particular focus on generative AI systems; (2) designing principled algorithmic frameworks for human-centered machine learning that model human preferences and behaviors, integrating these models into machine learning pipelines for applications such as healthcare, recommender systems, and education; and (3) evaluating and auditing the societal impacts of large-scale AI systems, including large language models and recommender systems. She graduated from the Machine Learning Department at Carnegie Mellon University in 2023 where she was advised by Zachary Lipton, and spent a year at Princeton Language & Intelligence as a postdoc. She has also spent time at Apple and Google DeepMind London during her Ph.D., and was an Open Philanthropy AI Fellow.

He He [homepage]
Assistant Professor, New York University

Title: Tracing LLM Capabilities to the Training Data

Abstract: Pre-trained large language models (LLMs) exhibit remarkable emergent capabilities. However, the origins of these capabilities remain poorly understood. What kind of patterns in pre-training data enable LLMs to perform ICL? Can LLMs infer causal relations from relational data in text, or are they limited to memorizing explicit causal facts? We explore these questions and discuss the broader implications for understanding how LLMs generalize from training data to perform complex tasks and where their reasoning capabilities face limitations.

About the speaker: He He is an Assistant Professor of Computer Science and Center for Data Science at New York University. Her current research focuses on understanding and aligning large language models, and human-AI collaboration. Before joining NYU, she obtained her PhD in 2016 from the University of Maryland, College Park, did a post-doc at Stanford, and spent one year at AWS. 

Gopala Anumanchipalli [homepage]
Assistant Professor, University of California Berkeley

Title: Grounding Speech AI models in Human Speech Mechanisms

Abstract: In this talk I will go over our recent attempts to induce human mechanisms in Spoken language AI applications. I will first go over the neural and physiological basis of fluent speech production, and talk about grounding current Self-Supervised Speech models (Hubert, WavLM etc) in human processes. Specifically, I will talk about probing articulatory information in these models. I will then present a new coding scheme called SPARC (SPeech ARticulatory Coding), that completely describes speech in terms of speech articulation, and discuss the universality of such a coding scheme across speakers and languages. I will also detail newer experiments in inducing higher order phonological structure into these models. As a practical demonstration of these developments, I will talk about some experiments in Neurotechnologies, specifically Brain-Computer Interfaces that use some of these results. Time permitting, I will briefly talk about our recent work in Dysfluent speech modeling characterization in speech disorders, exemplar based speech stylization etc. 

About the speaker: Gopala Anumachipalli is the Robert E. And Beverly A. Brooks Assistant Professor in the EECS department at UC Berkeley, where he leads the Berkeley Speech Group. He holds an adjunct position at UCSF, and is a member of Berkeley AI Research (BAIR), and Computational Precision Health (CPH). His group focuses on the science and engineering of spoken language, with application to human health — both for screening speech disorders and externally restoring lost function using Brain Computer Interfaces. He obtained his PhD from Carnegie Mellon University and went to UCSF for postdoctoral training. He has been recognized as a Kavli Fellow, Noyce Innovator,  Hellman Fellow, Google Research Scholar, JP Morgan AI Research awardee, among other honors.

Speaker Website:  https://people.eecs.berkeley.edu/~gopala/