蜜桃传媒

Motivated by observations of phenomenal cooperative behavior in nature, large teams of simple robots promise unprecedented task efficiency and resilience benefits over sophisticated standalone systems. To advance this paradigm of large multi-robot teams, aka robot swarms, we have developed a suite of algorithms that guide the collective behavior of multiple ground or aerial robots for operations such as signal source localization, multi-location response, area coverage and boundary mapping. These algorithms are built using the machinery of optimization, graph theory and metaheuristics. To demonstrate their potential impact, we have explored their application in simulated use cases such as flood response, finding skiers trapped under avalanche and offshore oil spill mapping. More recently, we have broken new ground on scalability of such autonomous robot swarm operations by innovating graph learning approaches to generate real-time policies for multi-robot task allocation and capacitated vehicle routing. A culmination of our work in swarm robotics has been the demonstration of new reinforcement learning frameworks and gaming environments (for human-swarm interaction) to model situation-adaptive tactical decisions for teams of UAVs and UGVs performing victim search in complex urban environments.

Modern robotics and mixed reality is driven by the ability of our devices/robots to perceive the environment using cameras and depth sensors. Localization and Mapping is a fundamental task in such reasoning. While research in Visual Simultaneous Localization and Mapping (VSLAM) has been underway for over a decade, it is still challenging to efficiently run these pipelines on modern embedded systems. I will discuss Edge-SLAM, a novel Visual SLAM pipeline that splits the computing between the mobile device and an edge server to efficiently use resources. I will also discuss other systems challenges in VSLAM with potential for future research. 

In this talk, I will cover two difficulties encountered when translating AI from research to deployment settings. 

The first concerns application of AI to medical imaging settings, specifically dermatology. Previous studies of artificial intelligence (AI) applied to dermatology have shown AI to have higher diagnostic classification accuracy than expert dermatologists; however, these studies did not adequately assess clinically realistic scenarios, such as how AI systems behave when presented with images of disease categories that are not included in the training dataset or images drawn from statistical distributions with significant shifts from training distributions. We aimed to simulate these real-world scenarios and evaluate the effects of image source institution, diagnoses outside of the training set, and other image artifacts on classification accuracy, with the goal of informing clinicians and regulatory agencies about safety and real-world accuracy. We have identified specific deficiencies and safety issues in AI diagnostic systems for skin cancer that should be addressed in future diagnostic evaluation protocols to improve safety and reliability in clinical practice.

The second concerns deployment of few-shot learning methods. Recent progress on few-shot learning largely relies on annotated data for meta-learning: base classes sampled from the same domain as the novel classes. However, in many applications, collecting data for meta-learning is infeasible or impossible. This leads to the cross-domain few-shot learning problem, where there is a large shift between base and novel class domains. While investigations of the cross-domain few-shot scenario exist, these works are limited to natural images that still contain a high degree of visual similarity. Here, we propose the Broader Study of Cross-Domain Few-Shot Learning (BSCD-FSL) benchmark, consisting of image data from a diverse assortment of image acquisition methods. This includes natural images, such as crop disease images, but additionally those that present with an increasing dissimilarity to natural images, such as satellite images, dermatology images, and radiology images. Extensive experiments on the proposed benchmark are performed to evaluate state-of-art meta-learning approaches, transfer learning approaches, and newer methods for cross-domain few-shot learning.

Recent years have witnessed an unexpected and astonishing rise of AI-synthesized fake media (aka DeepFakes), thanks to the rapid advancement of technology and the omnipresence of social media. Together with other forms of online disinformation, the DeepFakes are eroding our trust in online information and have already caused real damage (thinking about the recent fake video of President Zelensky during the Ukrainian War). It is thus important to develop countermeasures to limit the negative impacts of DeepFakes. In this presentation, I will first overview the types and making of DeepFake media. I will then highlight recent technical developments to fight DeepFakes using AI technology. Detection technologies and state-of-the-art are covered. I will introduce a DeepFake detection platform, DeepFake-o-meter, which is a collaboration with Realty Defender Inc. to assist users to authenticate DeepFakes. In addition, I will briefly introduce the more recent developments of preemptive techniques to protect the users from becoming victims of DeepFake attacks. I will also discuss the current drawbacks and limitations of these counter technologies, and the future of DeepFakes and forensics. This talk aims to help law enforcement officers and media forensic practitioners to understand the capacity of current counter-technologies to DeepFakes to better employ them in practice.

Public opinion polling’s reputation took a sizable hit in the wake of the 2016 elections, as statewide polls in key battleground states ended up understating popular support for Donald Trump. Forecasting models that relied heavily on national polls similarly missed the mark. In response to these apparent shortcomings, the American Association for Public Opinion Research (AAPOR) conducted a thorough post-mortem review of what went wrong with the polls in 2016. In this talk, I provide an overview of the AAPOR report and situate its findings within the context of a similar such effort written in response to another (in)famous polling disaster—the 1936 Literary Digest Poll. I also take this opportunity to provide a preliminary analysis of nonresponse bias attributable to differential levels of distrust in the media—a potential contributor to polling errors in 2016 that went largely overlooked by the AAPOR report. Finally, I briefly touch upon issues related to poll aggregation of the type popularized by sites like FiveThirtyEight and (in an earlier period) Pollster.com, incorporating this discussion into a broader conversation of the limits of big data approaches to measuring public opinion. 

In recent times, deep learning techniques for training artificial neural networks have been highly successful and transformative in a wide range of AI applications. But such techniques are “data hungry”, requiring large amounts of data to learn the substantial number of parameters in the architectures. But in our research, we have very limited data, hence, we propose intelligent training methods to circumvent this limitation.

We will explore the notion of generating new samples of parent-infant pairs based on the probabilistic deep-learning models trained on actual dyads. Such a model can be useful for anonymizing existing data and for extending our extant 60-dyad dataset. Furthermore, such a generative model can be developed more extensively within the context of causal reasoning methods, to provide more explainable predictions than what can be obtained from previous approach. This work benefits from the interplay between the interdisciplinary researchers involved, including investigators from developmental + quantitative psychology, behavioral science + nonverbal communication, and affective computing + machine learning.

Large neural networks excel in many domains, but they are expensive to train and fine-tune. A popular approach to reduce their compute or memory requirements is to replace dense weight matrices with structured ones (e.g., sparse, low-rank, Fourier transform). These methods have not seen widespread adoption (1) in end-to-end training due to unfavorable efficiency--quality tradeoffs, and (2) in dense-to-sparse fine-tuning due to lack of tractable algorithms to approximate a given dense weight matrix. To address these issues, we propose a class of matrices (Monarch) that is hardware-efficient (they are parameterized as products of two block-diagonal matrices for better hardware utilization) and expressive (they can represent many commonly used transforms). Surprisingly, the problem of approximating a dense weight matrix with a Monarch matrix, though nonconvex, has an analytical optimal solution. These properties of Monarch matrices unlock new ways to train and fine-tune sparse and dense models.

It is critical to assess the risks of AI used in high-stakes enterprise workflows along several dimensions beyond accuracy. In this hands-on session, you will learn to build trust in AI along two important dimensions – fairness and explainability - using two open-source libraries, AI Fairness 360 and AI Explainability 360.

The tutorial will first introduce various types of unwanted bias and algorithmic fairness. Then, using samples of working code (Jupyter notebooks), participants will be led through the process of evaluating, as well as mitigating, bias in AI models using a variety of fairness metrics and mitigation methods. We will then explore various techniques for generating explanations from AI models, and look at multiple scenarios, and associated code samples, for explaining model predictions in these scenarios.

Cell migration is critical for many vital processes, such as wound healing, as well as harmful processes, like cancer metastasis. Recent experiments of cells migrating in 3D fibrous matrices reminiscent of biological tissue show some cells experience large shape changes and are thought to utilize cytoplasmic streaming and intracellular pressure in generating leading-edge protrusions. It has been hypothesized that the adhesion of the cell to its external environment does not play an important role in this type of motility in contrast to cell migration on flat surfaces. It is not well understood how cells generate forces to facilitate migration in using this amoeboid mode. In this talk, dynamic computational models of single-cell migration are presented in a fibrous extracellular matrix and in a confined channel. Results show the non-trivial relationship between cell rheology and its external environment during migration.

Genomic sequencing has the power to trace pathogen movement through time and space by monitoring changes in the viral genome.  The SARS-CoV-2 pandemic is the first time this approach has been deployed on such a vast, global scale.  As a result, we have been able to track not only movement of this virus globally, but also its evolution in real time.  Mutagenesis of this virus through millions of infections has resulted in variants of concern that have increased infectivity and/or pathogenicity of SARS-CoV-2.  Early in the SAR-CoV-2 pandemic, we established a whole genome sequencing pipeline to assess lineages circulating in Western New York (WNY), in collaboration with Erie County Department of Health (ECDOH).  This effort evolved into a collaborative network of regional partners performing SARS-CoV-2 PCR testing, including Erie County Department of Health (ECDOH), Kaleida Health, Catholic Health, Erie County Medical Center, Cattaraugus County Department of Health and KSL Diagnostics. We are also now part of a New York State SARS-CoV-2 sequencing consortium.  Initial sequences revealed entry into the region via Europe, similar to observations in New York City (NYC).  However, as the pandemic progressed and variants of concern emerged, we observed distinct patterns in lineages relative to NYC and other parts of the state.  For example, B.1.427 (epsilon variant of concern) became dominant in WNY, before it was displaced by B.1.1.7 (alpha variant of concern), while epsilon was never dominant in NYC.  Our analysis of alpha, delta and omicron lineages in WNY indicated unique lineage patterns, multiple introductions into WNY and community spread. Through early identification of variants of concern, we were able to inform public policy through collaboration with ECDOH.  Our work highlights the importance of widespread, regional surveillance of SARS-CoV-2 across the United States.  More recently, we have implemented wastewater sequencing and have been able to monitor transitions from delta to omicron and BA.1 to BA.2 variants in Erie County.  This regional network has expanded our ability to monitor viral pathogens beyond SARS-CoV-2 and potentially identify novel pathogens.

Three research projects are highlighted in this talk that reflect the application of mathematical modeling approaches to the development of complex drug formulations administered in the female urogenital tract for HIV prevention, the optimization of population pharmacokinetic model structures using evolutionary algorithms and the identification of parameter estimates in nonlinear dynamic systems.

The talk introduces two research projects at the Center for Unified Biometrics and Sensors, one in the area of Document Processing and the other in Biometrics.  The first project is on lecture video summarization.  MOOCS feature large numbers of whiteboard and blackboard based lecture videos.  These videos are indexed primarily by a few keywords and search and retrieval ability is limited. Deep indexing of the videos facilitated by detection of semantic content such as equations, text and diagrams, enables richer search and retrieval applications. The second project is on feature aggregation strategies for fusion.   A majority of the modern methods used for template aggregation of set-based face recognition systems rely on learning to quantify the quality of images present in a template. While focusing on weighting the feature embedding based on this quality factor, they have overlooked aggregation strategies that can adapt the template's features to the paired template involved in matching.  We propose a template feature aggregation strategy that tailors a template's image set to mirror the properties exhibited by the target template. The proposed method produces state-of-the-art results on standard unconstrained face recognition datasets validating the advantages of such an aggregation strategy.

Although there are numerous positive aspects to the use of social media and online games, there are also risks, including cyberbullying. Up to 35-40% of adolescents report experiencing cyberbullying (i.e., willful and repeated harm inflicted through digital media devices) and there has been an increase in rates of cyberbullying in recent years. Artificial Intelligence (AI) is a promising approach to cyberbullying prevention/intervention and there has been increased interest in using AI (i.e., machine learning) to detect, prevent, and intervene with cyberbullying and other forms of online harassment among youth. This session will provide an overview of the Alberti Center for Bullying Abuse Prevention, relevant research, and limitations and future directions for research on the use of AI in cyberbullying prevention/intervention. 

This talk will overview some recent advances in Bayesian and semi-Bayesian learning methods for somatic mutation data-driven cancer-type characterization problems. Leveraging diverse statistical theories developed in the fields of computational ecology, natural language processing, and statistical signal processing, we have recently developed a multilevel statistical modeling framework that can aid principled probabilistic inference in these cancer characterization problems. Our model aims to address pertinent open problems in clinical research aiming to predict the cancer types of cells based on their genomic footprints, and natural scientific research aiming to estimate the association between genomic alterations and various cancer types. The model is implemented leveraging modern computational Bayesian methods. Our recent publications have evinced impressive predictive accuracy and remarkable predictor attributability of the model when implemented on large-scale publicly available databases.  

For more than 25 years, I have been a member of the PETSc Team developing a set of community numerical libraries for scientific computing (). I will discuss the advantages of this kind of organization, including generating research ideas, career advancement, and developing new research collaborations.

At Scalable Computing Research Group, we are researching how to scale challenging computational problems to modern high performance parallel computers. Since many problems in Machine Learning and Artificial Intelligence have non-trivial computational and data management requirements, they naturally fall into our orbit of interest. In this brief talk, I will highlight our work on efficient algorithms for Bayesian networks learning, and I will show how the need to prototype complex ML algorithms led to the development of new programming model for expressing optimization problems.

With the rise of social media, cyberharassment (e.g., cyberbullying and online hate) has become more prevalent in daily interactions and has thus been identified as a critical social problem.  It often involves inappropriate online behaviors and deliberate cyber threats against individuals,  or specific social groups on the grounds of characteristics such as race, sexual orientation, gender, or religious affiliation.  Although Machine Learning (ML) has immense potential for automatic cyberharassment detection and researchers are increasingly using ML techniques to address this important social problem, they face key challenges to effectively address cyberharassment using ML. In this talk, I will present our recent work on defending against cyberharassment using ML techniques. I will especially discuss (1)understanding and defending against visual cyberbullying; (2) detecting and explaining online hateful content, and (3) exploring the robustness of multimodal models employed in cyberharassment defense.

Storytelling is one of the oldest forms of human communication and a fundamental mechanism for information sharing, sense-making, and social transformation. Entertainment-education is a globally recognized strategy that leverages the power of storytelling purposefully incorporated in entertainment programming for social and behavior change. In this presentation, I will use three examples – one from the United States, one from India, and one from China – to illustrate the potential and challenges of using emerging technologies such as streaming platforms, social media, and AI chatbots to tell compelling stories, create transformative experiences, and facilitate positive change.

Medical scanners are widely used in screening and diagnosis in order to develop a personalized treatment plan for patients. Current patient examination workflow with medical scanners comprises several critical pre-scan events that involve manual operations and physical interactions between patients and medical professionals such as directing patients to the examination room, assisting them to pose according to scanning protocols, and positioning them for the scan, which may increase t This talk will cover several computer vision algorithms and systems we developed to make medical scanners more intelligent and efficient by automating the scanning procedure, as well as real-world examples of their applications.