Bluetalks Seminar Series

Modern information systems typically use data with heterogeneous models and data schemas. For example, a smart transportation system uses data generated from various sources, such as mobile devices, airborne sensing systems, traffic cameras, microphones, RFID readers, etc.

A single database for storing distinct data doesn't work. Running ETL (Extract-Transform-Load) processes, manual curation, and maintenance to have a single database is costly and requires significant effort. Therefore, data from these scenarios generally resides in the most appropriate data storage systems for their storage and access, such as relational databases, NoSQL, HDFS, processing frameworks, hybrid multimodal databases, or hybrid NewSQL systems.

The objective of this presentation is to introduce the main concepts about heterogeneous data storage systems, illustrating a solution in a use case scenario in the Oil and Gas area. The presentation includes the following topics: characterizing existing storage solution classes, presenting examples of systems; presenting a taxonomy for federated data systems, their requirements, and challenges; illustrating an implementation in a use case scenario.

The use case scenario includes activities for pre-processing geological data to generate data for training and validating Deep Learning (DL) models in the Oil and Gas area. The solution for this scenario will be illustrated using the PostgresSQL FDW (Foreign Data Wrapper). This solution allows creating tables in PostgresSQL that bring externally stored data from heterogeneous data storage systems.

About the speaker: Leonardo G. Azevedo has been Research Scientist within IBM Research Brazil since 2013. He is Ph.D. (2005) and MSc. (2001) from PESC/COPPE/UFRJ and Bachelor in Informatics from UFRJ (Rio de Janeiro-Brazil). He was a professor within UniRio (2006 to 2018) and a researcher of the Graduate Program in Informatics (PPGI/UniRio) (2009 to 2018). He has more than 20 years of experience in system development and applied research, working on projects for national and international organizations. His research areas are Distributed Sytems, Service-Oriented Architecture (SOA), Microservices Architecture (MSA), Databases, Provenance, Data Integration, Polystore, Knowledge Engineering, Ontologies, and Business Process Management (BPM).

The presentation is structured as an introduction to the topic, its relevance for organizations interested in using generative AI applications, and an exploration of solutions and challenges. The target audience is AI application managers and developers, who already have a basic understanding of what generative AI is (like ChatGPT, Dall-E, etc.), in both large organizations and startups. The agenda is structured with the following topics:

• Introduction: Generative AI and applications
• The need for value alignment
• Basic value alignment techniques
• Success and failure examples
• Technical challenges and the issue of fragility
• Conclusion

About the speaker: Claudio Pinhanez is a scientist, innovator, and professor. He has been with IBM Research since 1999, and today leads research in Conversational Intelligence in the laboratory of IBM Research in Brazil. He is also the Deputy Director of the C4AI, the Center for Artificial Intelligence created by a partnership of University of São Paulo, IBM, and FAPESP. Claudio got his PhD from the MIT Media Lab in 1999. He is an expert in artificial intelligence, human-machine interaction, conversational systems, and service science. He has more than 120 papers published in journals and scientific conferences, and more than 30 patents issued in the USA, Europe, and Japan.

The methods of machine learning and deep learning have been extensively explored in understanding the chaotic behavior of the atmosphere and promoting weather forecasting.

With the success of generative artificial intelligence (AI) using pre-trained transformer models for language and vision modeling through prompt engineering and fine-tuning, we are now moving towards general AI (AGI). Recent approaches using transformer models, machine learning incorporating physical constraints, and graph neural networks have demonstrated state-of-the-art performance in relatively narrow spatial-temporal scales and specific tasks.

Despite this progress, we are still in the early stages of developing a general AI model for regional climate models and meteorological models.

In this talk, Daniel Civitarese will review current AI approaches, primarily from transformer learning and operator literature in the context of meteorology.

We will provide our perspective on success criteria for basic weather and climate prediction models. We will also discuss how such models can compete effectively in tasks such as downscaling (super-resolution), identifying conditions leading to forest fire occurrences, and predicting subsequent meteorological phenomena across various spatial-temporal scales, such as hurricanes and atmospheric rivers.

In particular, we examine current AI methodologies and assert that they have matured sufficiently to design and implement a foundational meteorological model.

About the speaker: Daniel Salles Civitarese joined IBM Research in 2015 as a postdoctoral researcher and became a research scientist in 2017.​ As a post-doc, he developed NLP and visual comprehension tasks for the SYNAPSE neuromorphic chip. As a research scientist, he became one of the key contributors to deep learning for subsurface characterization and seismic interpretation. The group has developed new neural network architectures to process 3D subsurface data through his leadership. He has achieved state-of-the-art results for seismic facies analysis and published several papers at industry and AI conferences.​ In 2019, Daniel received the IBM Outstanding Technical Achievement because of his contributions. Since 2020, he investigates the use of geospatial AI techniques in climate.

Climate change has been driven by two gases: carbon dioxide and methane. Despite being less prevalent, methane has a significantly higher heat-trapping capacity per molecule than carbon dioxide, making it a key target for reducing global atmospheric warming. We will introduce how the use of remote sensing, combined with deep learning models, aids in monitoring vast areas for methane plume detection. This enables the remote identification of emitting sources and remediation of leaks, for example, from oil and gas exploration installations.

About the speaker: Maciel Zortea is a Research Scientist at IBM Research - Brazil. He has experience in the development of image analysis and machine learning methodologies. Maciel works in multidisciplinary projects involving the science and application of remote sensing to solve problems related to environmental monitoring. He obtained his doctorate from the Università degli Studi di Genova in 2007.​

In this presentation, we will discuss a study on how we are using artificial intelligence to understand urban growth. The ability to predict how cities will grow will have numerous positive implications for urban development, transportation, energy planning, and climate impact preparation. We will show how deep learning can be applied with vast Earth observation, population, and economic datasets with global coverage. Furthermore, we will demonstrate how large geospatial foundation models can be efficiently fine-tuned to update urban area data up to the present day, enabling the use of urban growth prediction models on a global scale.

About the speakers:

Maysa Macedo is a research scientist and diversity and inclusion board member at IBM Research. She is PhD in Computer Science at Institute of Mathematics and Statistics - University of Sao Paulo. Her thesis involves the development of a computational method for vessel tracking in medical imaging. At IBM Research she has developed applications in agriculture with satellite and drone images, natural resources with seismic images, future of work with forecasts using curriculum data and climate change with climate data.​ Her research interests include machine learning, computer vision, fairness and medical imaging analysis. She has contributed to the academic society by being a reviewer for several journals and being a member of scientific conference committees.​

Davi Misturini is a PhD candidate in Applied Mathematics at the State University of Campinas (UNICAMP), where he studies generalization theory in Deep Learning. He participates in the IBM Research internship program under the IBM Sustainability Accelerator for clean energies, through the implementation of machine learning algorithms based on spatio-temporal geospatial data.

With the success of Foundation models in the context of natural language processing, the scientific community has been investigating whether the same approach can be applied in other domains and contribute to various applications, including tasks related to climate emergency. Recently, IBM and NASA developed a Foundation model for remote sensing data, available through the HuggingFace platform. The model was trained with a large amount of satellite data processed by NASA and applied to remote sensing tasks, such as identifying flooded areas or areas devastated by fires.

During the talk, we will discuss Foundation models, focusing on the model developed by IBM and NASA. We will also present the open-source tool developed by IBM to adapt this type of model for applications that use remote sensing data.

About the speaker: Daniela Szwarcman graduated in Electronic Engineering from PUC-Rio, Rio de Janeiro, Brazil, in 2013. The following year, she joined the Semiconductor Laboratory at PUC-Rio and enrolled in the Master's program in Electrical Engineering, focusing on Nanotechnology. In 2016, she began her PhD in Electrical Engineering at PUC-Rio, but in the field of Artificial Intelligence. Her research topic was deep neural networks and Neural Architecture Search. In 2017, she started a research internship at IBM Research. She participated in numerous AI projects, developing and implementing deep models for problems involving spatial and temporal data. Daniela completed her PhD in 2020 and now works as a researcher at IBM Research, primarily developing and applying deep networks in the context of climate-related problems. Her research interests include deep learning, generative models, and physics-informed machine learning.