AI Innovation Institute

Abstract: Large Language Models (LLMs) have transitioned from standalone prediction interfaces into integrated systems that incorporate content protection, external knowledge retrieval, and multi-step reasoning. While these functional layers expand model capabilities, they also introduce complex, inter-component dependencies that create novel and systemic security risks. This research provides a systematic deconstruction of the structural vulnerabilities emerging across these functional layers.

In this proposal, we evaluate the security boundaries of LLM systems through three pivotal dimensions:
The Content Layer: We present Watermark under Fire, revealing the inherent fragility of content-based tracing mechanisms under adaptive perturbations and highlighting the limitations of surface-level safety measures.
The Retrieval Layer: We introduce GraphRAG under Fire to examine the security of topology-aware knowledge integration. We reveal how graph-based indexing can be exploited as a structural lever for high-success poisoning attacks.
The Reasoning Layer: We detail AutoRAN, the first framework demonstrating the hijacking of internal safety reasoning in Large Reasoning Models (LRMs). This work proves that the transparency of the reasoning process itself creates a critical and exploitable attack surface.

Collectively, these studies demonstrate a systemic failure of add-on safety mechanisms in securing the broader LLM ecosystem. By identifying recurring patterns of exploitation across different system layers, this research provides the necessary foundation for transitioning from reactive patching to a more unified and architecturally-grounded approach to AI trustworthiness.

Speaker: Jiacheng Liang

Zoom: https://stonybrook.zoom.us/j/6669990420?pwd=dkY0eEw5YXpPSWo3RUE4OE1oVW90UT09&omn=97367037382
Meeting ID: 666 999 0420
Passcode: 075299

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

https://stonybrook.zoom.us/j/91775729097pwd=Qlc5Nks0NmlyKzJwMjR0S0hrdVZ3QT09

Meeting ID: 917 7572 9097
Passcode: 555459


Abstract: As the saying goes, there are many ways to skin a cat.
While we don't want to go around skinning cats, the world of
optimization is rich with different problems, problem formulations,
and methods and approaches, each with different guarantees and
computational benefits. In this talk we will take a tour down the
problem of structured sparsity in sensing to see how one simple
problem can inspire a wide range of analysis and tools. First, I will
present the optimality conditions for a generalized structured sparse
problem, which can be geometrically visualized as alignment of vectors
and matrices. Then I will introduce three approximation methods for
the problem of phase retrieval, which are a twist on stochastic
gradient and coordinate descent methods. These methods leverage
fundamental numerical linear algebra concepts to give fast approximate
solutions to large-scale problems, which then after postprocessing can
produce more reliable sensing results.

Bio: Yifan Sun received her PhD in Electrical Engineering from the
University of California Los Angeles in 2015, with research focusing
on convex optimization and semidefinite programming. She was then
Technicolor Research and Innovation, focusing on machine learning and
data science applications. More recently, she completed two postdocs,
at the University of British Columbia in Vancouver, Canada and
L'Institut National de Recherche en Informatique et Automatique
(INRIA) in Paris, France.

Nam Nguyen

4-5pm, Dec 17 2020

https://stonybrook.zoom.us/j/94214254415?pwd=K1VoQml4cFdlVW51VW41dWtid2tJdz09



The molecular mechanisms and functions in complex biological systems
currently remain elusive. Recent high-throughput techniques, such as
next-generation sequencing, have generated a wide variety of
multiomics datasets that enable the identification of biological
functions and mechanisms via multiple facets. However, integrating
these large-scale multiomics data and discovering functional insights
are, nevertheless, challenging tasks. To address these challenges,
machine learning has been broadly applied to analyze multiomics. In
particular, multiview learning is more effective than previous
integrative methods for learning data's heterogeneity and revealing
cross-talk patterns. Although it has been applied to various contexts,
such as computer vision and speech recognition, multiview learning has
not yet been widely applied to biological data--specifically,
multiomics data. Therefore, we have developed a framework called
multiview empirical risk minimization (MV-ERM) for unifying multiview
learning methods (Nguyen, et al., PLoS Computational Biology, 2020).
MV-ERM enables potential applications to understand multiomics
including genomics, transcriptomics, and epigenomics, in an aim to
discover the functional and mechanistic interpretations across omics.
Based on MV-ERM, we have developed the following methods:
ManiNetCluster, Varmole and ECMarker.



(1) ManiNetCluster (Nguyen, et al., BMC Genomics, 2019) is a manifold
learning method which simultaneously aligns and clusters gene networks
(e.g., co-expression) to systematically reveal the links of genomic
function between different phenotypes. Specifically, ManiNetCluster
employs manifold alignment to uncover and match local and non-linear
structures among networks, and identifies cross-network functional
links. We demonstrated that ManiNetCluster better aligns the
orthologous genes from their developmental expression profiles across
model organisms than state-of-the-art methods. This indicates the
potential non-linear interactions of evolutionarily conserved genes
across species in development. Furthermore, we applied ManiNetCluster
to time series transcriptome data measured in the green alga
Chlamydomonas reinhardtii to discover the genomic functions linking
various metabolic processes between the light and dark periods of a
diurnally cycling culture;



(2) Varmole (Nguyen, et al., Bioinformatics, 2020) is an interpretable
deep learning method that simultaneously reveals genomic functions and
mechanisms while predicting phenotype from genotype. In particular,
Varmole embeds multi-omic networks into a deep neural network
architecture and prioritizes variants, genes and regulatory linkages
via biological drop-connect without needing prior feature selections.
With an application to schizophonia, we demonstrate that Varmole
provides an effective alternative for recent statistical methods that
associate functional omic data (e.g. gene expression) with genotype
and phenotype and that link variants to individual genes in population
studies such as genome-wide association study;



(3) ECMarker (Jin*, Nguyen*, et al., Bioinformatics, 2020) is an
interpretable and scalable machine learning model that predicts gene
expression biomarkers for disease phenotypes and simultaneously
reveals underlying regulatory mechanisms. Particularly, ECMarker is
built on the integration of semi- and discriminative- restricted
Boltzmann machines, a neural network model for classification allowing
lateral connections at the input gene layer. With application to the
gene expression data of non-small cell lung cancer (NSCLC) patients,
we found that ECMarker not only achieved a relatively high accuracy
for predicting cancer stages but also identified the biomarker genes
and gene networks implying the regulatory mechanisms in lung cancer
development.



Finally, we propose a novel multiview learning method, Malignomics, to
predict phenotypes from heterogeneous multi-omic features. Malignomics
will first align multi-omic features by deep manifold alignment onto a
common latent space, better predicting nonlinear relationships across
omics. This deep alignment aims to preserve both global consistency
and local smoothness across omics and reveal higher-order nonlinear
interactions (i.e., manifolds) among cross-omic features. Second, it
uses these manifold structures to regularize the classifiers for
predicting phenotypes. This manifold-regularization allows
highlighting cross-omic feature manifolds and prioritizing the
features and interactions for the phenotypes. The prioritized
multi-omic features will further reveal underlying phenotypic
functions and mechanisms and thus enhance the biological
interpretation of Malignomics. We will apply Malignomics to
multi-omics data in neuropsychiatric disorders, and prioritize gene
regulatory networks linking risk variants, regulatory elements, and
genes for the disorders. We will also compare Malignomics with the
state-of-the-arts, and investigate how the manifold regulation will
potentially improve understanding of multi-omics functions and
predicting diseases.

Title: AI-Driven Target Selection Methods for Touch and Gaze Input

Abstract: Accurately selecting targets is an essential aspect of  Human-Computer Interaction. Erroneous selections can cause tedious undo and redo actions. Additionally, some selection errors are non-reversible and can lead to undesirable consequences. However, high-accuracy target selection remains a challenge on touchscreen devices due to the small target size and imprecise touch inputs, and in gaze interaction because of the gaze tracking noise and no easy-to-use selection action. We first propose ReLM, a Reinforcement Learning-based Method for touchscreen target selection. ReLM can automatically show suggestions and require a second touch if the input is ambiguous, and can directly select a target candidate when the input is certain. Our empirical evaluation shows that ReLM reduces the error rate from 6.92% to 1.63%, and the selection time from 2.23s to 1.59s over Shift, an existing suggestion-based method. Compared to BayesianCommand, a direct selection-based method, our ReLM reduces the error rate from 3.64% to 0.89%, while increasing the selection time by only 200 ms. Secondly, we investigate how to improve target selection performance for gaze interaction. We propose BayesGaze, an eye-gaze based target selection method. It accumulates the signal of each gaze point for selecting a target calculated by Bayes Theorem, and uses a threshold mechanism to determine the target selection. Our investigation shows that BayesGaze improves target selection accuracy and speed over a dwell-based selection method, and the Center of Gravity Mapping method.

All are welcome. Here  is the zoom meeting link:
https://stonybrook.zoom.us/j/93130953411?pwd=Rm5IRlVPQ3M0cHJsTXpCVFljUlFGUT09Meeting ID: 931 3095 3411Passcode: 999413

Join Stony Brook University's Center for Excellence in Learning and Teaching (CELT) for a bootcamp on how to use AI to enhance your teaching and courses. This event will demonstrate how ChatGPT, Microsoft Copilot, 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.

Register here.

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.us/j/98775312249?pwd=uORNAnSdcssrPZdqOsqaMAF5aLcRD9.1
ID: 98775312249
Passcode: 505777