Shallow Learning

Learning objectives

• Explain the role of shallow learning in modern AI systems
• Implement key supervised algorithms (regression, classification, trees, ensembles, SVM) in Python using standard libraries
• Apply clustering and PCA to explore and structure data
• Describe and apply basic NLP workflows (tokenization, vectorization, embeddings) and compare them with shallow text-classification approaches.
• Design and run small ML experiments, including train/validation splits and model comparison
• Evaluate models with appropriate metrics and communicate their performance and limitations
• Plan and execute an ML mini-project following CRISP-DM and present the results clearly

Learning outcomes

• Build and tune shallow ML models for classification and regression problems
• Use unsupervised learning to discover structure in unlabeled data
• Build simple NLP pipelines using classic methods and embeddings, and explain when shallow or neural approaches are more suitable for text tasks
• Choose between alternative algorithms and justify their choice based on data and constraints
• Structure an end-to-end ML project from data preparation to evaluation and reporting
• Communicate results, limitations and next steps to technical and non-technical stakeholders

Week 1 – Supervised learning I (≈14 h)

• On-site sync (Welcome, dynamics, networking)
• Short video lectures: intro to ML, supervised learning, regression vs classification, trees.
• Readings & recommended videos to deepen the same topics
• Exercises: 2 real-life datasets (one classification, one regression) solved with multiple algorithms and compared
• Forum discussion and short quiz

Week 2 – Supervised learning II & Unsupervised learning (≈13 h)

• Video lectures on advanced supervised learning techniques (boosting, SVM) and clustering (k-means, DBSCAN, hierarchical)
• Readings & recommended videos to deepen the same topics
• Exercises on real datasets
• Forum discussion and quiz
• Live Q&A: troubleshooting, networking

Week 3 – Dimensionality reduction & NLP (context) (≈13 h)

• Video lectures: exploratory data analysis, PCA, UMAP, NLP, embeddings
• Readings and curated internet videos on the topics
• Practical “real life” problem using methods from weeks 1–3
• Forum discussion, quiz
• Live Q&A: troubleshooting, networking

Week 4 – Neural networks (≈13 h)

• Video lectures: basic neural networks, contrast with shallow methods
• Readings and curated internet videos on neural nets
• Practical “real life” problem
• Forum discussion, quiz
• Live Q&A: troubleshooting, networking

Week 5 – MLOps & Start Project (≈13 h)

• Video lectures: CRISP-DM, project management for ML, deployment and maintenance of models
• Individual or small-team final project (end-to-end ML workflow)
• Forum: participants present results via short slides or 5-min recorded talk
• On-site farewell session: recap, final Q&A, networking session

Week 6 – Project & Conclusion (≈13 h)

• Individual or small-team final project (end-to-end ML workflow)
• Participants present results via short slides or 5-min recorded talk
• On-site farewell session: recap, final Q&A, networking session

Carles Fenollosa

Caroline König

CARLES FENOLLOSA. Lecturer of artificial intelligence at the Polytechnic University of Catalonia–BarcelonaTech (UPC), professional educator, and private consultant. Computer engineer with a Master’s Degree in Artificial Intelligence from the UPC. He has been a researcher at the Barcelona Supercomputing Center, founder and CEO of several start-ups in the field of AI, and inventor of a patent. Author of “La singularidad” (Arpa Editores, 2024).

CAROLINE KÖNIG. Lecturer of artificial intelligence at the Polytechnic University of Catalonia- BarcelonaTech (UPC) and researcher of the IDEAI-UPC Research Center (Intelligent Data Science and Artificial Intelligence Research Center).  Computer engineer, Master in Advanced Methods in Artificial Intelligence and PhD in Artificial Intelligence by UPC. Over 10-years of experience in software development and research in the area of machine learning. Currently she is the scientific coordinator from UPC in the European research project Permepsy for the development of a personalized AI-based medicine platform.

Modern AI is often associated with deep learning, but a large fraction of real-world problems can be solved effectively with “shallow” models that are simpler, cheaper and easier to interpret. This course focuses on these methods and on how to use them responsibly in applied projects.

Over four weeks, participants review the main supervised and unsupervised techniques (linear and logistic regression, trees, ensembles, SVMs, clustering and PCA) and learn how to select, train and evaluate them on realistic datasets. A short module on deep learning and NLP provides context and a clear comparison point, helping participants understand when shallow or deep approaches are more appropriate.

The final week is devoted to a guided project where participants follow the CRISP-DM methodology from business understanding to deployment considerations. Along the way they document decisions, evaluate trade-offs and present their results to peers, gaining experience that can be transferred to their own research or industrial use cases.

<ul>
<li>Comfortable programming in Python (variables, functions, basic libraries)</li>
<li>Basic knowledge of linear algebra (vectors, matrices)</li>
<li>Introductory understanding of probability and statistics</li>
<li>Ability to work with CSV/tabular data and Jupyter notebooks is highly recommended</li>
</ul>

<p>Certificate of Achievement</p>

<ul>
<li>James et al., An Introduction to Statistical Learning. <a href="https://www.statlearning.com/"><u>https://www.statlearning.com</u></a></li>
<li>Russel and Norvig, Artificial Intelligence: A Modern Approach <a href="https://aima.cs.berkeley.edu/"><u>https://aima.cs.berkeley.edu</u></a></li>
<li>Scikit-learn user guide: sections on supervised learning, unsupervised learning and model evaluation</li>
</ul>

<ul>
<li>Computer with a minimum of 8 GB of RAM. 16GB recommended. A discrete GPU is not required, but can speed the model training process up.</li>
<li>Stable internet connection with the ability to video conference</li>
</ul>