Repository logo
 
Loading...
Project Logo
Research Project

LASIGE - Extreme Computing

Authors

Publications

Evolving robust and readable land cover models for improved forest monitoring
Publication . Batista, João E.; Silva, Sara Guilherme Oliveira da
A importância das florestas a nível mundial está bem estabelecida na sociedade, sabendo-se que a desflorestação tem sérias implicações na biodiversidade, nas sociedades rurais dependentes das florestas para comida e sustento, e nas emissões de gases de efeito de estufa que responsáveis pelas alterações climáticas. Mecanismos como o REDD+, desenvolvidos pela UNFCCC, ajudam países em desenvolvimento através de recompensas por evitarem a desflorestação e degradação florestal enquanto promovem a conservação das florestas, desenvolvimento sustentável e melhorias no sequestro de carbono nas florestas. No entanto, para estes mecanismos serem eficazes, é necessário monitorizar florestas de forma eficiente. Este trabalho contribui para os domínios científicos de ciências da computação e de ciências ambientais ao aplicar aprendizagem automática evolutiva neste tópicos, lidando com dois aspetos importantes da monitorização florestal: o mapeamento de tipos de coberto e a deteção de degradação florestal. Este trabalho também contribui para a área de aprendizagem automática através do estudo de métodos para engenharia de atributos que criem modelos robustos e interpretáveis que possam ser lidos, percebidos e corrigidos pelos especialistas das respetivas áreas científicas. Várias aplicações nossas conseguiram aumentar a robustez e transferibilidade de modelos de aprendizagem automática, melhorando o grau de confiança dos modelos de monitorização florestal. Estes modelos podem ser validados estudando se os atributos usados fazem sentido no contexto do problema, e também reduzindo a dimensionalidade dos conjuntos de dados ao ponto de se conseguir visualmente detetar problemas nos dados. Também estudamos a relação entre a complexidade de modelos com a sua interpretabilidade. Interpretabilidade e explicabilidade de modelos têm vindo a tornar-se tópicos populares nos últimos anos. Isto motiva trabalho futuro no uso de aprendizagem automática evolucionaria para forçar a evolução de modelos que sejam mais facilmente interpretáveis através de funções de fitness específicas.
Cooperative Autonomous Driving in Simulation
Publication . Costa, Gonçalo Manuel Valente da; Costa, António Casimiro Ferreira da; Cecílio, José Manuel da Silva
In this master’s thesis, we defined, implemented, and evaluated an architecture tailored to a simulation environment to assess cooperative autonomous driving protocols. The primary objective of this thesis is to showcase the benefits of employing this architecture to enhance driver and passenger safety in scenarios with coordination problems between vehicles. Additionally, this thesis aims to assess the resilience of the protocol implemented in the architecture. We tested it in demanding scenarios, such as intersections with a high volume of vehicles or situations with significant latency and communication failures between vehicles. These stress tests allow us to evaluate the protocol’s robustness and ability to prevent failures in such challenging conditions. Consequently, this thesis contributes by defining an architecture that facilitates vehicle-tovehicle communication for resolving coordination issues. This solution allows the incorporation of protocols specifically designed to address coordination challenges among vehicles. Furthermore, this thesis also makes a substantial contribution by implementing and testing the proposed architecture, incorporating a protocol responsible for coordinating vehicles at intersections to ensure the safety of drivers and passengers. To conduct the solution testing, we utilized a highly realistic simulator known as SVL. This simulator was crucial in creating the testing scenarios, including environmental conditions and vehicle dynamics. The key findings from our testing demonstrate that adopting this architecture mitigates accidents and potentially hazardous situations at intersections compared to scenarios where it is not utilized, resulting in improved safety. We also verified that the protocol can manage many vehicles, even when the communication latency between the vehicles is high.
Safe and resource-aware verification for programming Cyber-Physical Systems
Publication . Mão De Ferro, Carlos; Martins, Francisco Cipriano da Cunha; Costa, António Casimiro Ferreira da
Cyber-physical systems (CPS) integrate computation with physical processes and are expected to influence nearly 30 billion devices globally by 2030. Developing software for these systems presents unique challenges, especially in critical and resource-constrained scenarios. A predominant issue is the integration of multiple stateful systems, where the incorrect sequence of function calls compromises the overall system functionality. Formal verification increases safety, but its application is often confined to experts. This thesis presents Shelley, a novel model checking framework that streamlines the verification of function call order when integrating stateful systems. Shelley uniquely allows developers to specify temporal requirements closely aligned with the source code and automates the verification process. The framework’s ability to generate behavior diagrams and show intuitive counterexamples further assists developers in analyzing the program behavior, ensuring that the software is safe and adheres to its specified requirements. An evaluation is provided, which includes rigorous case studies such as NASA’s PHALANX experiment and an industrial setting, demonstrating the framework’s broad applicability and effectiveness in identifying invalid behaviors. Our contributions through the Shelley framework include the formalization of the process of extracting models from MicroPython code, as well as a new specification language tailored for verifying call order. In addition, we establish the decidability of checking system usage via automata theory and we introduce an automated process for checking requirements that utilizes temporal logic over finite traces. Together, these advancements simplify the incorporation of formal verification into high-level programming contexts, ensuring that more programmers can deliver safer software.
Deep learning system for biomedical relation extraction combining external sources of knowledge
Publication . Sousa, Diana; Couto, Francisco José Moreira
A Extração de Relações (ER) biomédicas bem-sucedida pode fornecer evidências aos investigadores sobre possíveis associações desconhecidas entre entidades, avançando o nosso conhecimento atual sobre essas entidades e os seus processos inerentes. As soluções atuais estado-da-arte para realizar ER biomédicas são baseadas em abordagens de aprendizagem profunda com arquiteturas compostas de múltiplas representações de dados, como derivados do BERT (p.e., BioBERT, PubMedBERT e SciBERT). No entanto, estes falham em recorrer a conhecimento externo para aumentarem o seu desempenho e tendem a utilizar apenas os dados de treino. O objetivo final deste projeto foi desenvolver um sistema de ER de alto desempenho que combine as representações de linguagem anteriores com conhecimento obtido de fontes externas, como ontologias específicas de domínio. O corpo principal deste trabalho apresenta três sistemas de aprendizagem profunda baseados em arquiteturas distintas e com diferentes abordagens à injeção de conhecimento, a saber, BiLSTMs, modelos de recomendação e representações de linguagem baseadas em BERT, todos integrados com informação proveniente de ontologias biomédicas (p.e., Gene Ontology e Human Phenotype Ontology). Esses sistemas superam o estado da arte anterior em ER biomédicas em conjuntos de dados amplamente utilizados como o DDI Corpus (interações medicamentosas) e o BC5CDR Corpus (interações entre compostos químicos e doenças). Esta tese também apresenta uma nova abordagem para a produção de conjuntos de dados de ER, utilizando técnicas de supervisão distante aliadas a plataformas de crowdsourcing para validação, resultando no PGR-crowd Corpus que descreve relações fenótipo humano-gene. Os sistemas e abordagens criados nesta tese foram aplicados e avaliados com sucesso em vários estudos (p.e., workshops, desafios e outras aplicações relevantes), por exemplo, ao serem premiados com a 7a posição no NASA LitCoin NLP Challenge de cerca de 200 equipas participantes e contribuindo para o esforço de investigação relativo ao COVID-19.
Deep semantic entity linking
Publication . RUAS, PEDRO; Couto, Francisco José Moreira
Knowledge organization systems, such as ontologies and knowledge graphs, are essential for organizing biomedical and clinical information and data. However, the growing volume of available scientific literature raises challenges in their maintenance. Entity linking approaches assist humans in curation by mapping entities described in text to entries of the knowledge organization systems, but their lack of coverage originates unlinkable or NIL entities. Besides, the state-of-the-art depends on deep learning models trained on large amounts of human-annotated data, which is hard to acquire. The present research work focuses on tackling these limitations of human-annotated data in the biomedical entity linking task. First, it addresses the lack of coverage of biomedical knowledge organization systems by using relation extraction to find missing relations and focusing on the problem of the NIL entities. Relation extraction increases the semantic information available for graph-based entity linking approaches (REEL), and focusing on the partial mapping of NIL entities (i.e. NIL entity linking) also improves the performance of such approaches (NILINKER). Second, the research work proposes a new deep learning model trained on a large-scale training dataset generated through automatic methods. The model is part of the pipeline X-Linker integrating different entity linking models, providing more flexibility and performance. The pipeline achieved state-of-the-art performance in the biomedical entity linking task in several datasets (BC5CDR-Disease, BioRED-Chemical, NCBI Disease). The described approaches and several others focusing on related tasks, such as named entity recognition, text classification, and recommendation of biomedical entities, were applied to several case studies, including competitions, workshops and challenges.

Organizational Units

Description

Keywords

Contributors

Funders

Funding agency

Fundação para a Ciência e a Tecnologia

Funding programme

6817 - DCRRNI ID

Funding Award Number

UIDP/00408/2020

ID