Aurélie Cools

Teaching Assistant - UMONS

Biography
Having graduated with a degree in Management Engineering from the renowned Louvain School of Management (LSM), I began my career journey where finance meets technology, serving as a Credit Analyst in a bank. In this role, I delved into financial analytics and creditworthiness evaluation, applying my theoretical knowledge to practical financial scenarios. Yet, my insatiable appetite for knowledge and my fascination with the intersection of management and technology led me further in my educational journey. In 2021, I earned a Master's degree in Civil Engineering, specializing in Computer Science and Management, further bridging my expertise in managerial and technological domains. My journey did not stop at the exploration of the financial sector; it extended into the academic and research realms. In September 2021, I embarked on a Ph.D. in Engineering Sciences, diving deeper into the myriad applications of technology. Simultaneously, I embraced the role of Contract Assistant in the Computer Science, Software, and Artificial Intelligence service at UMONS. This role not only supports my ongoing doctoral research but also serves as a platform to intertwine research with practical applications in the rapidly evolving field of artificial intelligence. With a firm belief in the symbiosis between theoretical insights and practical applications, I apply my blend of engineering management and computer science expertise with the vision of making significant contributions to both the financial and technological sectors.
Thesis

Towards Less Dependent Deep Learning Models in Terms of Input Data, Labels and Output Descriptors

ABSTRACT :

The current era is witnessing a substantial increase in data availability, attributed partly to continually expanding storage capacities and partly to the enhanced diversity of data sources. Both individuals and various professional sectors, including industry and medicine, generate myriad types of data daily, such as images from production lines, temperature curves, medical imaging, and more. This plentiful data availability propels the utilization of artificial intelligence and, notably, neural networks, across a wide range of fields, including unexpected ones like football, where AI is employed for player segmentation, ball tracking, match statistics, and more.

However, certain domains face data scarcity, while others encounter challenges with the lack of data annotations, which are costly. Given that the best-performing neural networks, trained in a supervised manner, require a substantial amount of annotated data, this presents various challenges. Furthermore, supervised training often generates large-sized descriptors, leading to extended calculation times and potential storage issues. This work seeks solutions to reduce the necessary quantity of data and annotations while maintaining a descriptor of minimal dimension.

KEYWORDS : Self supervised learning, few shot learning, data annotations, reduction dimension

Advisors

Sidi Ahmed Mahmoudi

Embedded and Explainable AI

Check Profile
Publications
A comparative study of reduction methods applied on a convolutional neural network

Electronics 11 (9), 1422, 2022

Access publication
A New Comparative Study of Dimensionality Reduction Methods in Large-Scale Image Retrieval

Big Data and Cognitive Computing 6 (2), 54, 2022

Access publication
CARENET: A NOVEL ARCHITECTURE FOR LOW DATA REGIME MIXING CONVOLUTIONS AND ATTENTION

Access publication