An Introduction to Deep Domain Adaptation with Emphasis on Remote Sensing Applications
Gilson Costa - State University of Rio de Janeiro
Abstract: The presentation starts by introducing some formal definitions of Transfer Learning and Domain Adaptation. Next, the fundamentals of two of the main Domain Adaptation approaches used for Computer Vision problems, namely Representation Matching and Appearance Adaptation, will be presented. In sequence some seminal Domain Adaptation methods will be described, intertwined with solutions for exemplary applications, mostly related with Remote Sensing problems.
Short-bio: He obtained the title of Doctor in Electrical Engineering from the Pontifical Catholic University of Rio de Janeiro (PUC-Rio) in 2009, having carried out part of his doctoral research at the Leibniz University Hannover, Germany. He graduated in Computer Engineering (PUC-Rio) in 1991. He obtained a Master's degree in Computer Engineering, with emphasis on Geomatics, at the State University of Rio de Janeiro (UERJ), in 2003. He carried out post-doctoral research between 2010 and 2015 at the Department of Electrical Engineering at PUC-Rio, and between 2019 and 2020 at the Institute for Photogrammetry and Geoinformation (IPI) at Leibniz University Hannover. He is an Associate Professor at the State University of Rio de Janeiro (UERJ), Department of Informatics and Computer Science, and the current Coordinator of the Postgraduate Program in Computational Sciences and Mathematical Modeling (PPG-CompMat) at the UERJ’s Institute of Mathematics and Statistics (IME/UERJ). He has more than 100 articles published in specialized journals, book chapters and annals of national and international conferences. He participated in various research and innovation projects, including several international cooperation projects. He is currently carrying out research in the areas of Image Analysis and Computer Vision, primarily in Remote Sensing applications, such as detecting deforestation in tropical forests. He is a member of the International Society of Photogrammetry and Remote Sensing (ISPRS), the Brazilian Chapter of the IEEE Geoscience and Remote Sensing Society (GRSS), and the Brazilian Chapter of the Society of Latin American Specialists in Remote Sensing (SELPER).
Transforming Healthcare: 3D Human Interaction in Telemedicine
Andrea Britto - Microsoft Research
Abstract: In this talk, Andrea will delve into the groundbreaking 3D Telemedicine project, an initiative aimed at increasing access to healthcare for rural and underserved communities in Africa through live 3D communication between patients and doctors. Utilizing Microsoft’s Holoportation technology and low-cost Azure Kinect sensors, this project provides a 360-degree view for remote clinical consultations, significantly enhancing patient satisfaction and the quality of medical evaluations. Andrea will also discuss the challenges and opportunities in using AI to improve the quality of 3D models, as well as for motion captioning and segmentation. By integrating these advancements, the project aims to further bridge the gap between in-person and remote consultations, ensuring that distance is no longer a barrier to quality healthcare.
Short-bio: Andrea is a Senior Research Software Engineer at Microsoft Research. She holds a B.Sc. and M.Sc. in Computer Science from IME-USP. Andrea's research interests include Computer Vision, Machine Learning, and Computer Graphics. She has extensive experience in the industry, holding more than 20 patents. At Microsoft Research, Andrea’s project portfolio includes innovative initiatives such as paralinguistic prompting and 3D Telemedicine, where she aims to enhance healthcare accessibility in underserved regions.
Prithvi WxC: Foundation Model for Weather and Climate
Daniel Civitarese - IBM Research
Abstract:The realization that AI emulators can rival the performance of traditional numerical weather prediction models on HPC systems has led to the development of numerous large AI models addressing use cases such as forecasting, downscaling, and nowcasting. While parallel developments in AI literature focus on foundation models—models that can be effectively tuned to address multiple, diverse use cases—developments in the weather and climate domain largely emphasize single-use cases, particularly mid-range forecasting. In this lecture we will cover what are foundation models, and how they are being developed to weather and climate. Next, we will discuss Prithvi-WxC, a 2.3-billion-parameter foundation model developed using 160 variables from the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2). We will cover how Prithvi-WxC encodes regional and global information, its training in a distributed environment to maximize GPU utilization and minimize memory footprint, and the downstream tasks that NASA and our team are experimenting with.
Short-bio: Daniel has worked as a research scientist at IBM Research Brazil since 2017, working in the Spatiotemporal Modeling group to advance the state-of-the-art in developing and evaluating spatiotemporal predictive models using data-driven machine learning approaches, such as deep learning and large language models a.k.a. foundation models.
Assessing timber trade networks and supply chains in Brazil
Luis G. Nonato - University of São Paulo
Abstract: Forest degradation in the Brazilian Amazon is driven by factors such as fire, mining, and illegal logging. The Brazilian government has implemented control mechanisms to combat illegal timber extraction that have positively impacted deforestation rates. Under these regulations, all wood products, from raw logs to processed lumber, must be registered in control systems before transportation. This allows for the analysis of wood products transported between companies over time. However, the existence of three partially integrated control systems complicates a full analysis of the timber market. In this talk, we will present a data science project focused on integrating data from these systems, aiming to create timber trade networks, which help identify companies or groups operating outside expected standards. We also propose a method to trace likely supply chains of timber companies, addressing long-standing government concerns about timber traceability. Among the results, we show that certain timber trade networks have components that operate without connections with licensed forests, suggesting that unregistered timber is input into those components, which is illegal. Additionally, we illustrate how supply chain analysis can considerably enhance customer confidence in the legality of purchased timber products.
Short-bio: Dr. Luis Gustavo Nonato is a Full Professor at the Institute of Mathematical and Computer Sciences at the University of São Paulo (USP) in São Carlos, Brazil. Dr. Nonato’s research interests encompass visualization, machine learning, and data science. Dr. Nonato has also been a visiting professor at the Center for Data Science at New York University and a visiting scholar at the Scientific Computing and Imaging Institute at the University of Utah.
Large Multimodal Models for Open Ended Generation and Understanding
Vicente Ordóñez - Rice University
Abstract:In this talk I will provide an overview of how the field of computer vision has been impacted by the recent success of large multimodal models and what are some of the opportunities to build more sophisticated models that can work for open ended tasks. Particularly, I will discuss some of our recent work on adapting multimodal models for open vocabulary visual grounding through limited extra annotated data, self-consistency regularization and synthetic data. I will also describe our work on adapting large multimodal models for fine-grained question answering, step-by-step reasoning and text-to-image generation. Fine-grained question answering requires optimizing for correctness in a setup where incorrect answers are very similar to correct answers. Our work demonstrates the importance of verifying and reinforcing correct answers both during training or inference. Text-to-image generation is a complementary capability that is typically optimized independently for models trained for image-to-text generation. Our work shows how some large pre-trained models can be used as composable modules to achieve a single framework for both generation and understanding. Finally, I will highlight some recent successes in scaling up both multimodal generative and understanding models and some current trends in this area.
Short-bio: Vicente Ordóñez-Román is an Associate Professor in the Department of Computer Science at Rice University and Visiting Academic at the Amazon AGI Foundations team. He is also affiliated with the Ken Kennedy Institute at Rice University. His research interests lie at the intersection of computer vision, natural language processing and machine learning. He is a recipient of a Best Paper Award at the conference on Empirical Methods in Natural Language Processing (EMNLP) in 2017 and the Best Paper Award -- Marr Prize at the International Conference on Computer Vision (ICCV) in 2013. He has also been the recipient of an NSF CAREER Award, an IBM Faculty Award, a Google Faculty Research Award, a Google Award for Inclusion Research and a Facebook Research Award. Vicente obtained his PhD from the University of North Carolina at Chapel Hill, and has also been a visiting researcher at the Allen Institute for Artificial Intelligence and a visiting professor at Adobe Research.
PaleoScan: Large-Volume Fossil Scanning
Claudio T. Silva - New York University
Abstract:Fossils are crucial for understanding our natural history and the digitalization of fossils has paved the way for paleontologists to share and study them in greater detail. Yet, many fossil-dense regions, in particular low- and middle-income countries, lack the resources to digitalize their vast collections. This project reports on a collaboration between paleontologists and computer scientists to design, build, and operate a device that can be deployed in the field for digitizing a collection of thousands of fossils. We introduce PaleoScan, a user-friendly, cost-effective, high-volume scanner designed to expedite the digitization of extensive fossil collections. PaleoScan is a self- contained 3D scanning system consisting of a light and compact mirrorless camera, a microcontroller, a ChArUco calibration board, and user-controlled LEDs. Software and data processing is cloud-based, where the user interacts with the system through a web application. We deployed PaleoScan at the Museum of Paleontology Placido Cidade Nuvens, a museum in Brazil with a world-class fossil collection. Our early results reveal its potential to revolutionize the scanning process for fossils. The PaleoScan project is a collaborative effort that includes researchers from NYU, Universidade Regional do Cariri (URCA), Brazil, Universidade Federal de Pernambuco, Brazil, Universidade Federal do Ceara, Brazil, New York Institute of Technology, and AMNH. The project has received funding from NYU, Fundação Cearense de Apoio ao Desenvolvimento (Funcap), and the National Science Foundation.
Short-bio: Cláudio T. Silva is Institute Professor of Computer Science and Engineering and Data Science at New York University. His research interests include visualization, visual analytics, machine learning, reproducibility and provenance, geometric computing, urban computing, computer graphics, and computer vision. He received his BS in mathematics from the Universidade Federal do Ceará (Brazil) in 1990, and his MS and PhD in computer science at the State University of New York at Stony Brook in 1996. Claudio has advised 20+ PhD, 10 MS students, and mentored 20+ post-doctoral associates. He has published over 300 publications, including 20 that have received best paper awards. He has over 27,500 citations according to Google Scholar. Claudio is active in service to the research community and is a past elected Chair of the IEEE Technical Committee on Visualization and Computer Graphics (2015–18). Claudio is a Fellow of the IEEE and has received the IEEE Visualization Technical Achievement Award. He was the senior technology consultant (2012-17) for MLB Advanced Media’s Statcast player tracking system, which received a 2018 Technology & Engineering Emmy Award from the National Academy of Television Arts & Sciences (NATAS). Our lab’s work has been covered in The New York Times, The Economist, ESPN, and other major news media.
Cohesion and Separability of Hierarchical Clusters via Linkage Methods
Eduardo Laber - PUC-Rio
Abstract: Clustering techniques, in general, consist of grouping objects so that similar objects are in the same group, while dissimilar objects are in different groups. These techniques are widely used in exploratory data analysis and to accelerate various computational tasks. Although they emerged a long time ago, there is still a lot of research in this area. Linkage methods are a very popular class of heuristics for obtaining hierarchical clusters. Although this class has been widely used in applications for decades, knowledge about its theoretical properties is still quite limited. In this talk I discuss a series of recent results that seek to characterize the cohesion and separability of classical linkage methods such as single-link, complete-link and average-link.
Short-bio: Eduardo Laber is an Associate Professor of the Computer Science department, at PUC-RIO. He was a visiting scholar at the University of California, San Diego (UCSD) from July 2023 to July 2024. Eduardo has supervised numerous master's and doctoral students and published papers in some of the most prestigious forums in the subfield of Algorithms and Theory, including STOC, SODA, SICOMP, and JACM. In 2015, he received the Best Paper Award at ISAAC 2015, the Asian Conference on Algorithms. In recent years, he has been developing new algorithms and analyses for problems relevant to machine learning. His results have been regularly published in ICLM and NeurIPS, flagship conferences in the field.
Challenges and Opportunities of Data Science in the Beverage Industry
João Gomes - Federal University of Ceará
Abstract: The beverage industry, like any other sector, faces complex and constantly evolving challenges. In this talk, we will explore how data science can be used to overcome these challenges and drive innovation. I will present practical cases that demonstrate the power of data analysis to optimize processes, improve decision-making and create new business opportunities.
Short-bio: Dr. João Paulo Pordeus Gomes is an Associate Professor in the Department of Computer Science at the Federal University of Ceará (UFC). He earned his bachelor’s degree in Electrical Engineering from UFC in 2004, followed by a master’s and doctorate from the Instituto Tecnológico de Aeronáutica in 2006 and 2011, respectively. Between 2006 and 2013, he served as a technological development engineer at EMBRAER S/A. His research interests encompass digital signal and image processing, pattern recognition, and fault detection and prognosis. As of February 2024, Dr. Gomes has been serving as the Head of Data & Analytics at Solar Coca-Cola in Fortaleza, Ceará, Brazil.
Spotlighting Key Moments in First-Person Videos by Video Acceleration
Erickson Nascimento - Federal University of Minas Gerais
Abstract:In this talk, we will explore several semantic fast-forward approaches designed to compress first-person Videos into shorter, more engaging versions. These methods prioritize maintaining the video's contextual integrity while emphasizing its most relevant segments to preserve the semantic richness of the original content. First-Person Videos are often characterized by monotonous, lengthy, and unedited footage captured by body-mounted devices, making them visually unappealing and tedious to watch. This creates a pressing need for solutions that enable quick access to critical information by accelerating playback and crafting a compressed, contextually coherent version of the video.
Short-bio: Erickson is an Associate Professor of Computer Science at the Federal University of Minas Gerais (UFMG), Brazil, with extensive research experience across both academia and industry. He was a visiting researcher at Berkeley AI Research (BAIR) at the University of California, Berkeley, and served as a Computer Vision Architect at Microsoft. His work, published in leading conferences and journals such as CVPR, ICCV, NeurIPS, ACCV, TPAMI, and IJCV, focuses on Computer Vision and Pattern Recognition, including multi-modal image and geometric data representation, image matching, and video processing for egocentric videos.
From Small Molecules to Side Effects: Machine Learning in Drug Discovery
Diego Galeano - National University of Asunción
Abstract: In this talk, we will explore how machine learning techniques, including the sChemNET framework and large language models (LLMs) architectures such as RAG and GraphRAG, can revolutionize drug discovery and medicine. From predicting small molecules that target microRNA bioactivity to accurately retrieving drug side effects in LLM systems, these technologies advance our ability to understand and predict complex biological interactions, paving the way for safer and more effective therapeutic solutions.
Short-bio: Dr. Diego Galeano is a Machine Learning Researcher at the Faculty of Engineering, Universidad Nacional de Asunción (FIUNA) in Paraguay. Dr. Galeano’s research focuses on developing machine-learning models for applications in drug discovery and medicine. He has contributed to studies on predicting drug side effects and developing ML frameworks for pharmacological countermeasures on the impact of space radiation on human health.
Integrating Data for the Diagnosis and Treatment of Mental Disorders
Deisy Gysi - Federal University of Paraná
Abstract:Many mental disorders share overlapping symptoms, making accurate diagnosis—and consequently effective treatment—a significant challenge. Currently, the gold standard for diagnosis relies on inventories administered during clinical consultations, with no molecular follow-up. Similarly, as the symptomatology of mental disorders is complex and symptoms are shared across different conditions, the genetic basis of these disorders also exhibits considerable overlap. In this lecture, I present a study employing a multi-omics approach, network-based analysis, and machine learning to identify biomarkers for the diagnosis of six neuropsychiatric disorders.
Short-bio: Deisy Gysi is a Professor of Statistics specializing in the development and application of computational and statistical methods to address complex problems in science and medicine. Her research spans areas such as machine learning, network science, bioinformatics, and personalized medicine, with a focus on advancing diagnostic tools and therapeutic discovery. Passionate about interdisciplinary collaboration, she actively contribute to bridging quantitative methodologies with real-world applications in mental health, genomics, and other fields.