Fall 2019 Commencement: https://www.stonybrook.edu/commcms/commencement/
Abstract:
Recent advances in deep learning have significantly enhanced the capabilities of Natural Language Processing (NLP) and Vision-Language Models (VLMs). However, these advancements come with increased vulnerabilities, notably through backdoor attacks that pose severe security threats. This thesis addresses two critical dimensions of Trustworthy AI and Efficient Multimodal Representation Learning: (1) security through analyzing, detecting, and designing backdoor attacks in NLP and VLMs, and (2) efficiency through advanced multimodal representation methods tailored for clinical and medical imaging applications.
In the first dimension, we explore the internal mechanisms exploited by backdoor attacks, identifying the distinctive phenomenon of attention focus drifting in compromised transformer models, where trigger tokens consistently hijack attention. Leveraging these insights, we propose robust detection frameworks, including the attention-based Trojan detector (AttenTD) and a task-agnostic logit-based detection method (TABDet), achieving effective identification of backdoored NLP models across diverse tasks. We further introduce novel backdoor attack methodologies: the Trojan Attention Loss (TAL), enhancing attack efficiency and stealth through direct attention manipulation, and BadCLM, demonstrating critical vulnerabilities in clinical decision-support systems by effectively compromising clinical language models.
Extending our security exploration to multimodal settings, we investigate backdoor attacks on Vision-Language Models (VLMs), particularly in complex image-to-text generation tasks, proposing innovative techniques (TrojVLM, VLOOD) capable of embedding backdoors without direct access to original training data, thus showcasing practical risks in real-world scenarios.
In the second dimension, we address efficiency and interpretability challenges in clinical and pathology applications. We introduce TCP-LLaVA, the first multimodal large language model (MLLM) designed explicitly for Whole Slide Image (WSI) Visual Question Answering (VQA). Utilizing a novel token compression mechanism inspired by transformer-based models, TCP-LLaVA substantially reduces computational resource consumption while maintaining superior VQA performance across multiple tumor subtypes. Additionally, we present a multimodal transformer model integrating structured Electronic Health Records (EHR) with clinical notes, demonstrating enhanced predictive accuracy and interpretability for in-hospital mortality prediction through integrated gradient-based interpretability methods.
Together, these contributions present a comprehensive approach to ensuring AI models are not only secure against malicious manipulation but also efficient and interpretable for critical clinical applications, underscoring the essential need for trustworthy and effective AI systems.
Speaker: Weimin Lyu
Zoom: https://stonybrook.zoom.us/j/
Meeting ID: 239 232 6575
Passcode: 436192
Recent advances in deep learning have significantly enhanced the capabilities of Natural Language Processing (NLP) and Vision-Language Models (VLMs). However, these advancements come with increased vulnerabilities, notably through backdoor attacks that pose severe security threats. This thesis addresses two critical dimensions of Trustworthy AI and Efficient Multimodal Representation Learning: (1) security through analyzing, detecting, and designing backdoor attacks in NLP and VLMs, and (2) efficiency through advanced multimodal representation methods tailored for clinical and medical imaging applications.
In the first dimension, we explore the internal mechanisms exploited by backdoor attacks, identifying the distinctive phenomenon of attention focus drifting in compromised transformer models, where trigger tokens consistently hijack attention. Leveraging these insights, we propose robust detection frameworks, including the attention-based Trojan detector (AttenTD) and a task-agnostic logit-based detection method (TABDet), achieving effective identification of backdoored NLP models across diverse tasks. We further introduce novel backdoor attack methodologies: the Trojan Attention Loss (TAL), enhancing attack efficiency and stealth through direct attention manipulation, and BadCLM, demonstrating critical vulnerabilities in clinical decision-support systems by effectively compromising clinical language models.
Extending our security exploration to multimodal settings, we investigate backdoor attacks on Vision-Language Models (VLMs), particularly in complex image-to-text generation tasks, proposing innovative techniques (TrojVLM, VLOOD) capable of embedding backdoors without direct access to original training data, thus showcasing practical risks in real-world scenarios.
In the second dimension, we address efficiency and interpretability challenges in clinical and pathology applications. We introduce TCP-LLaVA, the first multimodal large language model (MLLM) designed explicitly for Whole Slide Image (WSI) Visual Question Answering (VQA). Utilizing a novel token compression mechanism inspired by transformer-based models, TCP-LLaVA substantially reduces computational resource consumption while maintaining superior VQA performance across multiple tumor subtypes. Additionally, we present a multimodal transformer model integrating structured Electronic Health Records (EHR) with clinical notes, demonstrating enhanced predictive accuracy and interpretability for in-hospital mortality prediction through integrated gradient-based interpretability methods.
Together, these contributions present a comprehensive approach to ensuring AI models are not only secure against malicious manipulation but also efficient and interpretable for critical clinical applications, underscoring the essential need for trustworthy and effective AI systems.
Speaker: Weimin Lyu
Zoom: https://stonybrook.zoom.us/j/
Meeting ID: 239 232 6575
Passcode: 436192
You are cordially invited to attend the biweekly Brookhaven AI Mixer (BAM). BAM includes one short talk on AI research happening at BNL, followed by an open mixer over coffee and snacks for everyone to network and discuss all things AI. The first half hour will consist of presentations that will be available via ZOOM, and the second half hour will be for in person only networking.
Abstract: Two-dimensional (2D) materials such as graphene, hBN, and TMDs offer atomically sharp interfaces and unprecedented tunability when vertically assembled into van der Waals heterostructures. These stacks have enabled discoveries ranging from moiré superconductivity and correlated insulators to quantum emitters and next-generation nanoelectronic devices. Yet constructing high-quality heterostructures remains largely artisanal: researchers manually identify exfoliated flakes, align a polymer stamp by eye, and finely adjust temperature and contact geometry through tacit skill. This manual workflow is difficult to reproduce, scales poorly, and prevents systematic exploration of the enormous combinatorial space of materials, twist angles, and interfacial conditions. AutoLab is an autonomous platform that translates this tacit human expertise into programmable, feedback-driven control. Instead of pressing flakes with predefined trajectories, AutoLab uses machine vision to detect polymer-wafer contact, dynamically regulates contact evolution through closed-loop actuation and temperature control, and captures high-quality flakes with the cleanliness and precision of expert manual fabrication. The system integrates perception, decision making, and motion planning into a single robotic framework, enabling reproducible stacking, wafer-level coverage, and accelerated discovery. Beyond 2D materials, AutoLab illustrates a broader paradigm for AI-native scientific automation: codifying human experimental reasoning into algorithms that interrogate data in real time, adaptively adjust instrumentation, and generate scalable, high-fidelity datasets. Such platforms could generalize to diverse research domains--quantum device fabrication, optical alignment, surface science, autonomous microscopy, and other workflows where expert intuition currently limits throughput and reproducibility. By bridging artisanal manipulation and robotic autonomy, AutoLab points toward a future where scientific discovery is accelerated by machines that not only execute instructions, but learn, respond, and collaborate with human scientists.
Biography: Dr. Yutao Li is a research associate from Department of Condensed Matter Physics and Material Science, Brookhaven National Laboratory. He has 8 years of experience in 2D material sample fabrication, and investigation in their electronic transport, optical and mechanical properties.
Join us every other Tuesday at noon in CDSD's Training Room (building 725, 2nd floor) to learn about interesting AI methods and applications, engage with potential collaborators, prepare for pending FASST funding calls, and build a community of AI for Science at BNL.
Location: CDS, Bldg. 725, Training Room
Join ZoomGov Meeting: https://bnl.zoomgov.com/j/1604383624?pwd=ffQ5cUPNxTI7nzClKQO6cnsNbhF9Vf.1
Meeting ID: 160 438 3624
Passcode: 558449
Abstract: As intelligent systems become more integrated into human environments, fostering trustworthy human-AI collaboration presents a pressing challenge. In this talk, I examine the interplay between an agent's performance and social dynamics in shaping trust in human-AI interactions. My approach combines testbed development, behavioral prototyping, and user study design to create controlled experimental setups that capture real-world interaction complexities, such as ambiguity, multi-agent dynamics, and conflicting goals.
I illustrate this with a recent VR study on multi-user interaction with an autonomous vehicle (AV). Moving beyond dyadic interactions, the study probes human perspectives from the roles of a pedestrian, driver, and AV passenger, all interacting with the AV simultaneously at an ambiguous all-way stop sign intersection. We compare interactions with efficient and prosocial AV behavior strategies, revealing diverging trust perceptions and preferences across user roles. These insights inform a broader research trajectory focused on balancing performance with social considerations in designing trustworthy human-AI collaborations.
Bio: JiHyun Jeong is a postdoctoral researcher at Cornell University working on human-computer interaction and human-robot interaction. Her research develops prototypes and methods to explore performance and social factors that influence collaboration and trust between humans and artificial agents. She holds a Ph.D. and MPS in Information Science from Cornell University, and a BSc in Computer Science and Engineering from Korea University. She is a recipient of an honorable mention for best paper at DIS.
Zoom: https://stonybrook.zoom.
Meeting ID: 987 3823 4619
Passcode: 474618
Join Stony Brook University's Center for Excellence in Learning and Teaching (CELT) for a boot camp on how to use AI to enhance your teaching and courses. This event will demonstrate how ChatGPT, Microsoft Copilot, NotebookLM, and other generative AI platforms can support you in crafting learning objectives, writing exam questions, composing rubrics, and designing course content such as lesson plans, in-class activities, instructional videos, and more.
https://stonybrook.zoom.us/j/92511854285?pwd=QRTHfULqHMWxJYoVyt3piOhNxWLfvs.1
https://stonybrook.zoom.us/j/92511854285?pwd=QRTHfULqHMWxJYoVyt3piOhNxWLfvs.1
The overall purpose of this seminar is to bring together people with interests in Computer Vision theory and techniques and to examine current research issues. This course will be appropriate for people who already took a Computer Vision graduate course or already had research experience in Computer Vision.
To enroll in this course, you must either: (1) be in the PhD program or (2) receive permission from the instructors.
Each seminar will consist of multiple short talks (around 15 minutes) by multiple students. Students can register for 1 credit for CSE656. Registered students must attend and present a minimum of 2 talks. Everyone else is welcome to attend. Fill in https://forms.gle/q6UG9ygauLp2a8Po8 to subscribe to our mailing list for further announcement.
Abstract: Reward hacking, where a reasoning model exploits loopholes in a reward function to achieve high rewards without solving the intended task, poses a significant threat. This behavior may be explicit, i.e. verbalized in the model's chain-of-thought (CoT), or implicit, where the CoT appears benign thus bypasses CoT monitors. To detect implicit reward hacking, we propose TRACE (Truncated Reasoning AUC Evaluation). Our key observation is that hacking occurs when exploiting the loophole is easier than solving the actual task. This means that the model is using less effort than required to achieve high reward. TRACE quantifies effort by measuring how early a model's reasoning becomes sufficient to obtain the reward. We progressively truncate a model's CoT at various lengths, force the model to answer, and estimate the expected reward at each cut-off. A hacking model, which takes a shortcut, will achieve a high expected reward with only a small fraction of its CoT, yielding a large area under the reward-vs-length curve. TRACE achieves over 65% gains over our strongest 72B CoT monitor in math reasoning, and over 30% gains over a 32B monitor in coding. We further show that TRACE can discover unknown loopholes during training. Overall, TRACE offers a scalable unsupervised approach for oversight where current monitoring methods prove ineffective.
Speaker: Dikshya
Location: Old Computer Science Building - CS2311
Speaker: Dikshya
Location: Old Computer Science Building - CS2311
Abstract:
Capturing the spatio-temporal (4D) dynamics of humans has been a long standing research problem in computer vision and graphics. Synthesizing photorealistic human avatars has broad applications, ranging from immersive telepresence in AR/VR and the movie industry, to enriching the education and healthcare systems. Earlier approaches relied on hand-engineered models that use a small amount of data from one or more subjects. With the advent of neural networks, training on large datasets enhanced the output visual quality. Currently, the combination of neural networks with graphics techniques has achieved natural-looking human animation. However, most approaches are identity-specific, trained only on a single identity, and use only one modality.
In this thesis, we address the problem of learning neural representations of humans in a holistic way. Given that the video data in the real world include multiple modalities (audio and video) and multiple identities, we develop multi-modal and multi-identity representations. First, we propose to reconstruct the 4D face geometry of humans by leveraging both audio and video information. In this way, the network produces accurate lip shapes and is robust to cases when either modality is insufficient. Next, we introduce a NeRF-based representation for audio-driven human face animation that achieves high-quality lip synchronization for cinematic content. Since humans communicate with their full body, combining body pose, hand gestures, and facial expressions, we extend our network to capture the full-body human motion for multiple identities simultaneously. In order to better disentangle identity and non-identity specific information, we subsequently study non-linear interactions between latent factors of variation, and propose a specific multiplicative module. In this way, we learn a multi-identity NeRF that robustly animates human faces under novel expressions and achieves a significant decrease in the total training time. Similarly, we propose a multi-identity gaussian splatting representation for human bodies, by constructing a high-order tensor. Assuming a low-rank structure, we learn a tensor decomposition that leads to a significant decrease in the total number of learnable parameters, as well as to a robust animation under novel poses. In the future, we propose to jointly synthesize audio and visual outputs from just text input. Given the recent rise of large language models, coupling text with natural-looking avatars can enhance the overall interaction between a human and an AI system.
Speaker: Aggelina Chatziagapi
Where: NCS, Room 220
Zoom link: https://stonybrook.zoom.
ID: 98775312249
Passcode: 505777
Capturing the spatio-temporal (4D) dynamics of humans has been a long standing research problem in computer vision and graphics. Synthesizing photorealistic human avatars has broad applications, ranging from immersive telepresence in AR/VR and the movie industry, to enriching the education and healthcare systems. Earlier approaches relied on hand-engineered models that use a small amount of data from one or more subjects. With the advent of neural networks, training on large datasets enhanced the output visual quality. Currently, the combination of neural networks with graphics techniques has achieved natural-looking human animation. However, most approaches are identity-specific, trained only on a single identity, and use only one modality.
In this thesis, we address the problem of learning neural representations of humans in a holistic way. Given that the video data in the real world include multiple modalities (audio and video) and multiple identities, we develop multi-modal and multi-identity representations. First, we propose to reconstruct the 4D face geometry of humans by leveraging both audio and video information. In this way, the network produces accurate lip shapes and is robust to cases when either modality is insufficient. Next, we introduce a NeRF-based representation for audio-driven human face animation that achieves high-quality lip synchronization for cinematic content. Since humans communicate with their full body, combining body pose, hand gestures, and facial expressions, we extend our network to capture the full-body human motion for multiple identities simultaneously. In order to better disentangle identity and non-identity specific information, we subsequently study non-linear interactions between latent factors of variation, and propose a specific multiplicative module. In this way, we learn a multi-identity NeRF that robustly animates human faces under novel expressions and achieves a significant decrease in the total training time. Similarly, we propose a multi-identity gaussian splatting representation for human bodies, by constructing a high-order tensor. Assuming a low-rank structure, we learn a tensor decomposition that leads to a significant decrease in the total number of learnable parameters, as well as to a robust animation under novel poses. In the future, we propose to jointly synthesize audio and visual outputs from just text input. Given the recent rise of large language models, coupling text with natural-looking avatars can enhance the overall interaction between a human and an AI system.
Speaker: Aggelina Chatziagapi
Where: NCS, Room 220
Zoom link: https://stonybrook.zoom.
ID: 98775312249
Passcode: 505777