Speaker: Omid Asghari

Date: 10th of April, 2024

Time: 4:00pm

Place: Room G4.1

Presentation title: "Sensitivity Analysis of Safety Metrics for Monitoring UAV Operations in U-Space"

Short bio:

Omid is a third-year Ph.D. candidate at the University of Coimbra. He earned his Bachelor's degree in Computer Engineering - Software from the University of Kurdistan and his master's degree from the Islamic Azad University, Science and Research Branch in Tehran. During his master's program, he gained practical experience in the industry as a software developer and application security specialist for six years. Omid's research interests primarily focus on U-Space safety assessment and the integration of analytical safety assessments with experimentation.

Abstract:

In recent years, UAVs have increasingly been utilized in urban environments due to their agility in movement, mechanical simplicity, affordability, and capacity to access locations that are challenging or impossible for humans to reach. With a significant number of drones expected to operate in urban airspace soon, enhancing safety through monitoring drone operations in U-space is essential. To achieve this monitoring, several safety metrics need to be calculated as measurement units.

The goal of this research is to monitor drone operations in U-space and calculate UAV operation risks by conducting sensitivity analyses on various safety metrics. This involves assessing the impact of different parameters on these metrics.

__________________________________________________________________

Speaker: Jiawei Wang

Date: 10th of April, 2024

Time: 4:00pm

Place: Room G4.1

Presentation title: "AI-based Safety-critical Components"

Short bio:

Jiawei Wang is a Ph.D. student at CISUC, University of Coimbra. She received her master's degree in Software Engineering from Beijing Institute of Technology, China, in 2020, with a specialization in Machine Learning applied to perception tasks. Under the supervision of Prof. João Campos, her current research is centered on characterizing and improving safety of AI-based components by addressing biases between data in training and deployment phases.

Abstract:

Artificial intelligence (AI) has become indispensable in safety-critical applications because of its exceptional performance. However, the inherent "black-box" nature often leads to incidents resulting in loss of property and lives. While AI's capability to autonomously learn from big data surpasses traditional algorithms, the quality of the dataset sets the upper limit on model performance. Dataset bias has remained a persistent challenge in machine learning (ML) since its start. Contemporary approaches such as data augmentation offer some mitigation against bias effects. While achieving comparable performance on data distinct from the training set remains challenging. In our work, we aim to enhance AI safety by identifying, transferring, and mitigating dataset-related biases. In particular, we will consider AI used in the perception components. Our preliminary results reveal there exist distinct dataset-related patterns across various image datasets for pedestrian classification task. Our next step is designing experiments to overcome the influence caused by dataset bias using the ideas from Generative Adversarial Networks (GANs).