Deep Learning Projects from a Regional Council: An Experience Report

Jónathan Heras

Keywords: Regional data, Deep Learning, Assignments, Experience report

TL;DR: Lessons learned and challenges of using open regional data in deep learning projects in an Artificial Intelligence course

Abstract: Due to the impact of Deep Learning both in industry and academia, there is a growing demand of graduates with skills in this field, and Universities are starting to offer courses that include Deep Learning subjects. Hands-on assignments that teach students how to tackle Deep Learning tasks are an instrumental part of those courses. However, most Deep Learning assignments have two main drawbacks. First, they use either toy datasets, that are useful to teach concepts but whose solutions do not generalise to real problems, or employ datasets that require specialised knowledge to fully understand the problem. Secondly, most Deep Learning assignments are focused on training a model, and do not take into account other stages of the Deep Learning pipeline, such as data cleaning or model deployment. In this work, we present an experience in an Artificial Intelligence course where we have tackled the aforementioned drawbacks by using datasets from the regional council where our University is located. Namely, the students of the course have developed several computer vision and natural language processing projects; for instance, a news classifier or an application to colourise historical images. We share the workflow followed to organise this experience, several lessons that we have learned, and challenges that can be faced by other instructors that try to conduct a similar initiative.