This educational innovation project focuses on teaching how to train deep neural networks for image analysis, combining project-based learning with participation in competitive challenges. The initiative arises from the need to equip students with practical skills in artificial intelligence, particularly as modern models demand large datasets and powerful computational resources. However, current university infrastructure often limits access to adequate hardware, making it difficult for students to experiment with state-of-the-art models. This project aims to bridge that gap by providing an environment where students can develop skills in training and optimizing models, aligning their experience with industry standards.

The approach is grounded in project-based learning (Botella Nicolás et al., 2019), encouraging students to participate in real-world image analysis competitions on platforms such as Kaggle (Bojer & Meldgaard, 2021) and Grand-Challenges (Armato et al., 2023). This strategy enables students to tackle real problems, work in multidisciplinary teams, and use objective metrics to assess their performance. The process is structured in stages: students are given a problem to solve, provided with the necessary computational resources, and tasked with developing innovative solutions that are later evaluated based on their effectiveness. Learning success is measured not only by accuracy, but also by creativity and effort in the proposed solutions.

The results so far show significant progress in student learning. In the Advanced Computer Vision course, students participated in challenges such as inferring plant characteristics from images, using a dataset of over 62,000 files. However, the complexity of the problem, combined with limited computational resources, reduced the variety of solutions submitted. In another edition, students tackled the detection of invasive plant species in natural images, achieving accuracies up to 0.988, compared to 0.997 in the best public Kaggle solution. In the Medical Image Segmentation course, students developed models to segment vertebral structures in thoracic-lumbar CT scans, combining AI with manual post-processing techniques.

These results highlight the importance of having adequate computational infrastructure to foster a broader range of approaches in student solutions. Additionally, gamifying the learning process through participation in competitive challenges has proven to be an effective strategy to boost student motivation and engagement. Despite notable progress, challenges remain—such as ensuring equitable access to computational resources so that all students can experiment with advanced models.

The project has successfully integrated artificial intelligence into the educational environment, enabling students to acquire key skills for their professional development. The challenge-based methodology has proven effective in enhancing practical understanding of AI models while fostering creativity and critical thinking. Nonetheless, continued improvement in resource availability remains essential to maximize the initiative’s impact. This approach not only strengthens academic training but also prepares students to tackle real-world problems, meeting the demands of an ever-evolving job market.

Armato. DOI: 10.1259/BJR.20221152/7499015
Bojer. DOI: 10.1016/J.IJFORECAST.2020.07.007
Botella. DOI: https://www.scielo.org.mx/scielo.php?script=sci_arttext&pid=S0185-26982019000100127

Keywords: Artificial intelligence, gamification, project-based learning, competition.