This workshop is intended for researchers, practitioners, students, and industry professionals in AI, robotics, machine learning, human-robot interaction, and related fields.

Workshop Overview:

Instead of learning from data alone, an embodied AI system learns through its movements, sensors, and interactions with the environment. This form of active, experience-based learning, informed by ongoing self-evaluation of its own abilities, enables embodied AI systems to adapt on the fly, understand context rather than just commands, and collaborate with humans in more natural and trustworthy ways.

Workshop Goals:

  1. Foster interdisciplinary dialogue across AI, robotics, and cognitive science.
  2. Identify key challenges and future research directions in embodied intelligence.
  3. Examine the role of embodiment in advancing toward AGI.

This workshop is Invitation-only. Please email Dr. IV Ramakrishnan (ram@cs.stonybrook.edu) to attend.

Read the announcement: https://mcusercontent.com/237207911c0fd4c1f78dd8524/files/070dec2e-a2f5-143e-0fe2-c4ebecdb5193/Embodied_AI_Workshop_Invitation_.pdf

Abstract : Humans reason about everyday situations by making commonsense-based inferences, derived both from explicitly stated information and implicit, unstated knowledge. In this thesis, I investigate whether NLP models have different aspects of causal knowledge about events and how to improve their understanding of narratives and plans.
Answering questions about why people perform actions in a narrative can test whether NLP systems contain and can effectively apply causal knowledge about events. I introduce TellMeWhy, a dataset concerning why characters in short narratives perform the actions described. An evaluation of then SOTA finetuned models show that they are far worse than humans. To improve models, it is important to understand what aspects of causal knowledge they need and how to best use external sources to inject this knowledge. In KnowWhy, I analyze different ways of injecting knowledge into models, which is difficult since we do not know apriori what type of knowledge will be needed to answer a question, hence requiring a ranking model to pick the most important inference. Results show that this retrieved knowledge helps models of all sizes, thereby improving their understanding of narratives.
Next, I study whether models can reason about causal aspects of plans. I focus on testing whether they understand the underlying causal dependencies reflected in the temporal order of a plan's steps. I introduce CAT-Bench, and find that SOTA models are underwhelming, and that model answers are not consistent across questions about the same step pairs. In their current state, these models cannot yet reliably be used for complex user-facing tasks. I then measure contemporary models' ability to perform user-facing and user-centric plan customization. I introduce the use of semi-symbolic edits in large language model (LLM) based agents and test several multi-LLM-agent architectures for plan customization. While LLMs still lack the ability to understand complex customization hints, my results suggest that LLM-based architectures may be worth exploring further for other customization applications. Finally, I distill complex reasoning capabilities into small language models (SLMs) using synthetic data that reflects a decomposition-then-editing process for plan customization. I demonstrate that explicitly teaching this latent causal reasoning significantly improves the quality of SLM-generated customizations. Overall, my work has improved how well NLP models understand complex reasoning associated with events in different contexts.

Speaker: Yash Kumar Lal

Location: NCS 220 or Zoom https://stonybrook.zoom.us/j/95849648243?pwd=dgPpZtDpgwQrK9z1SaPpNbBifaorzk.1
Abstract: Recent studies have highlighted the vulnerability of Natural Language Processing (NLP) and Vision-Language Models (VLMs) to backdoor attacks, posing significant security risks. Understanding these attack strategies is crucial for assessing model robustness and developing effective defenses. This thesis proposal aims to investigate the vulnerability of language and vision-language models, analyze abnormal behaviors in backdoor-attacked models, and develop defense methods to enhance safety of modern machine learning models at deployment.


We investigate the internal mechanisms of backdoored NLP models, identifying a distinct attention focus drifting phenomenon, where trigger tokens hijack attention regardless of the input context. Through comprehensive qualitative and quantitative analysis, we provide insights into the underlying mechanisms that enable backdoor attacks. Building on these insights, we propose detection methods to differentiate backdoored models from clean ones, through inspecting both the attention distribution and the model predictions. To better understand the vulnerability, we develop advanced backdoor attack strategies targeting language models in classification tasks. For BERT variants, we introduce Trojan Attention Loss (TAL), a novel method that directly manipulates attention patterns to enhance backdoor effectiveness, ensuring stealth and robustness. Vision-Language Models have demonstrated strong performance in recent years. Yet their vulnerability is largely underexplored. We investigate advanced backdoor attack strategies on Vision-Language Models, focusing on image-to-text generation tasks. We demonstrate how backdoors can be embedded in complex multimodal tasks while maintaining semantic integrity under poisoned inputs. Additionally, we propose innovative techniques for injecting backdoors without requiring access to the original training data, expanding the feasibility of real-world attacks.

This proposal provides novel insights into the internal mechanisms of backdoored models, propose effective detection strategies, and develop advanced attack techniques that expose critical vulnerabilities. These findings underscore the urgent need for robust security measures to defend against emerging backdoor threats in deep learning models. The results have been published in top venues including ICLR, ECCV, NAACL, EMNLP, etc.

Speaker: Weimin Lyu


Zoom link: https://stonybrook.zoom.us/j/99880605139?pwd=cfWbRG6n9v3GXEa7OqvXa5cOp5eLBv.1
Meeting ID: 998 8060 5139
Passcode: 843302

Abstract:
Deep learning models have achieved remarkable success across a wide range of computer vision tasks, including image classification, semantic segmentation, etc. However, such success highly relies on a large amount of annotated data, which are expensive to obtain. Moreover, their performance often degrades when there exist distribution shifts between training and test data. Domain Adaptation overcomes these issues by transferring knowledge from a label-rich source domain to a related but different target domain. Despite its popularity, domain adaptation is still a challenging task, especially when the data distribution shifts are severe, while the target domain has no or few labeled data.

In this thesis, I develop four efficient domain adaptation approaches to improve model performance on the target domain. Firstly, inspired by the large-scale pretraining of Vision Transformers, I explore Transformer-based domain adaptation for stronger feature representation and design a safe training mechanism to avoid model collapse in the situation of a large domain gap. Secondly, I observe that source models have low confidences on the target data. To address this, I focus on the penultimate activations of target data and propose an adversarial training strategy to enhance model prediction confidences. Thirdly, I study using weak supervision from prior knowledge about target domain label distribution. A novel Knowledge-guided Unsupervised Domain Adaptation paradigm is devised, and a plug-in module is designed to rectify pseudo labels. Lastly, I step into the task of Active Domain Adaptation, where the labels of a small portion of target data can be inquired. I propose a novel active selection criterion based on the local context and devise a progressive augmentation module to better utilize queried target data. The robustness of domain adaptation approaches, in addition to accuracy, is critical yet under-explored. To conclude the thesis, I empirically study set prediction in domain adaptation using the tool of conformal prediction and conformal training.


Location: New Computer Science Bldg., Room 120
Zoom Link: https://stonybrook.zoom.us/j/92736258273?pwd=ipDdh1CTG6dRYmqa3ltUvooei8OfaT.1Meeting ID: 927 3625 8273
Passcode: 466399
Abstract: Many unresolved legal questions over LLMs and copyright center on memorization: whether specific training data have been encoded in the model's weights during training, and whether those memorized data can be extracted in the model's outputs. While many believe that LLMs do not memorize much of their training data, recent work shows that substantial amounts of copyrighted text can be extracted from open-weight models. However, it remains an open question if similar extraction is feasible for production LLMs, given the safety measures these systems implement. We investigate this question using a two-phase procedure: (1) an initial probe to test for extraction feasibility, which sometimes uses a Best-of-N (BoN) jailbreak, followed by (2) iterative continuation prompts to attempt to extract the book. We evaluate our procedure on four production LLMs -- Claude 3.7 Sonnet, GPT-4.1, Gemini 2.5 Pro, and Grok 3 -- and we measure extraction success with a score computed from a block-based approximation of longest common substring (nv-recall). With different per-LLM experimental configurations, we were able to extract varying amounts of text. For the Phase 1 probe, it was unnecessary to jailbreak Gemini 2.5 Pro and Grok 3 to extract text (e.g, nv-recall of 76.8% and 70.3%, respectively, for Harry Potter and the Sorcerer's Stone), while it was necessary for Claude 3.7 Sonnet and GPT-4.1. In some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim (e.g., nv-recall=95.8%). GPT-4.1 requires significantly more BoN attempts (e.g., 20X), and eventually refuses to continue (e.g., nv-recall=4.0%). Taken together, our work highlights that, even with model- and system-level safeguards, extraction of (in-copyright) training data remains a risk for production LLMs.

Speaker: Xinyue

Location: CS2311

Talk Title: Knowledge-enhanced LLMs and Human-AI Collaboration Frameworks for Creativity Support


Abstract:

Large language models (LLMs) constitute a paradigm shift in Natural Language Processing and Artificial Intelligence. To build AI systems that are human-centered, I propose we need knowledge-aware models and human-AI collaboration frameworks to help them solve tasks ultimately aligning these models better with human values. In this talk, I will discuss my research agenda for human-centered AI with a case study on creativity that focuses on how to augment LMs with external knowledge, build effective human-AI collaboration frameworks as well as theoretically grounded robust evaluation protocols for measuring capabilities of NLG systems. I will begin by describing knowledge-enhanced methods for creative text generation such as metaphors. Next, I will describe how content creators can collaborate and benefit from the creative capabilities of text-to-image-based AI models. Finally, I will focus on the design and development of theoretically grounded evaluation protocols to benchmark the creative capabilities of Large Language Models in both producing as well as assessing creative text. I will end this talk by highlighting the current limitations of existing models and future directions toward building better models that will enable efficient and trustworthy human-AI collaboration systems.


Bio:

Tuhin Chakrabarty is a final-year Ph.D. candidate in the Natural Language Processing group within the Computer Science department at Columbia University. His research is supported by the Columbia Center of Artificial Intelligence & Technology (CAIT) & an Amazon Science Ph.D. Fellowship. He was also a Computational Journalism fellow at NYTimes R&D and an intern at the Allen Institute of Artificial Intelligence, Salesforce Research, and Deepmind. His research interests are broadly in Natural Language Processing, Computer Vision, and Human-Computer Interaction with a special focus on Human-Centered Methods for Understanding, Generation, and Evaluation of Creativity. His work has been recognized at top natural language processing and human-computer interaction conferences and journals such as ACL, NAACL, EMNLP, TACL, and CHI. He has been involved in organizing several workshops and tutorials at NLP conferences such Figurative Language Processing workshop at EMNLP 2022, NAACL 2024, and the tutorial on Creative Text Generation at EMNLP 2023. His work on AI and creativity has been mentioned in mainstream news media such as The Hollywood Reporter and more recently The Washington Post.

Join Zoom Meeting https://stonybrook.zoom.us/j/97103601583?pwd=TnpGMXdpeEd1N0hZcXppS1BLNFJhZz09 (ID: 97103601583, passcode: 004031) Join by phone (US) +1 646-931-3860 (passcode: 004031) Joining instructions: https://www.google.com/url?q=https://applications.zoom.us/addon/invitation/detail?meetingUuid%3DILacj94mRvSXgTYt0Cqs1w%253D%253D%26signature%3D9f2f1e7e603bbcb9034724d084eea8846c19a38b7436180170dfc3f1d718b425%26v%3D1&sa=D&source=calendar&usg=AOvVaw3MsNgLSPMRl8L5i6BosYrB Meeting host: H.Andrew.Schwartz@stonybrook.edu

Join Zoom Meeting:
https://stonybrook.zoom.us/j/97103601583?pwd=TnpGMXdpeEd1N0hZcXppS1BLNFJhZz09
Title: Class visual similarity based noisy sample removal in generative Few Shot Learning
Time: Thursday, Feb 4, 11:30am - 1:00pm
Zoom:
https://stonybrook.zoom.us/j/8563646526?pwd=anJna1gzUStXNlNVSUIzdDRUSC9CUT09

Meeting ID: 856 364 6526
Passcode: 203791



Abstract:  

Over the past decade, larger datasets, hardware accelerations, and network architecture improvements have contributed to phenomenal achievements in many tasks of computer

vision. However, in the absence of large datasets, computer vision models struggle to learn

general representations which results in poor performance. Few-shot learning tries to address 

this problem by proposing models which learn from a few examples.


I first give an overall review of few-shot learning methods. I particularly focus on generative Few Shot Learning(FSL) methods, which augment the scarce categories in a dataset by generating samples for those rare categories. As the actual class distribution can be complex and lie very close to each other, the sample generated for one class can be noisy or lie close to another class.  However, none of the current FS generative methods perform any form of quality control of the generated samples.


In this work, I propose to identify and remove the generated samples that are less likely to be in the distribution of the few-shot class. Here I particularly deal with few-shot scenarios where the

prior information of the relationship between the classes based on visual  similarity is available. The main idea is to exploit these priors to better identify the unreliable generated samples.


Particularly, I have proposed two methods based on class relationship to detect noisy generated samples. In the first method, we assume that the embedding space of each class follows a Gaussian distribution.  From this assumption, I propose Gaussian Neighborhood (GN), a method to estimate how likely a generated sample is drawn from the estimated distribution of a few-shot class.  We evaluate this method on the Hematopoiesis dataset.  By simply eliminating samples based on thresholding our proposed GN scores, the few-shot  classification  performance  is  improved by 5% and 2% in five shot and one shot respectively, compared to the model trained on all generated images. 


The GN scores represent the similarity distances from the generated samples to their classes, based on the assumption that each class is a Gaussian distribution.  However, this assumption might be strict in many scenarios since the real distributions of data can be arbitrarily complex.  Thus in my second proposed method, I aim to learn such similarity distances directly from data via metric learning. I propose to train a deep-network to regress the similarity distance between a pair of samples.  This network is trained using both the class-level  visual  similarity  information  and  the  class  labels.   This method improves the 1-shot and 5-shot classification performances by 0.5% and 1% respectively, compared to GN.