The third introductory AI course, built around a private Kaggle competition: students design, compile and combine neural network models, and read research articles related to the competition’s goal.

Type

Course (part 3 of 3)

Duration

60h - 15 blocs/sittings

Competition

Private Kaggle competition for Collège Ahuntsic students

Technologies

Kaggle, Python, TensorFlow and Keras, Google Colab

Materials

4 slide decks (PDF)

Language

French

View the course on GitHub

What the course covers

The course opens on Kaggle: its notebooks and datasets, how to join the course’s private competition, how to submit predictions as a CSV file, and the competition rules. Changing labels by hand or reusing another team’s results counts as plagiarism, and mixing the training and test sets is treated as a major error.

The technical lectures then go deeper into models: sizing the inputs to the features (table columns or image pixels), interconnections, outputs, layers and architectures; advanced compilation, with loss functions for classification and regression (such as cross-entropy, MAE and MSE), optimizers and metrics, followed by training, evaluation and plotting; and finally ensemble methods: hard, soft and weighted voting, stacking, bagging and boosting.

Research in conference format

From session 7 to 12, the class works in a conference format on articles related to the competition’s topic. It is a first contact with reading papers, reviewing the literature and finding new ML knowledge on one’s own. Teams then present their work in the last two sessions.

Structure

• Session 1: Kaggle and competition rules
• Session 2: creating ML/DL models
• Session 3: advanced model compilation
• Session 4: ensemble models
• Session 5: work period
• Session 6: midterm exam
• Sessions 7–12: conference-format articles on the competition topic
• Sessions 14–15: presentations

View the course on GitHub