Advanced Applied Data Analytics

This course is the second, more advanced part of the diiP course “Applied Data Analytics.” The first part covers basic techniques, methodologies, and practical skills for data analysis. The second part delves into more advanced topics in Machine Learning and Deep Learning (see schedule below).

Like the first part, this course is intended for a variety of doctoral students from several faculties at Université Paris Cité. Key topics include data science, data analysis, machine learning, deep learning, and data mining.

Format of Lectures

This intensive course runs over one week in the winter, from Monday to Friday, between 10:00 AM and 1:00 PM, totalling approximately 15 hours.

Instruction includes slides, hands-on Google Colaboratory exercises, and questionnaires.

The course is conducted via Zoom.

Information

The last session took place from January 20-24, 2025. The next session will be held in January 2026.

Interested students can enrol through https://u-paris.fr/doctorat/advanced-applied-data-analytics/.

Enrolled students who need access to course materials should contact me via email.

Learning Outcomes

By the end of this course, students will:

• Understand the fundamentals of data-driven learning, including key AI and ML concepts.

• Gain familiarity with transformer architectures, their applications in sequence modelling, and how they power modern AI models.

• Develop an in-depth understanding of Graph Neural Networks (GNNs) and their use in processing structured data

• Learn about Self-Supervised Learning, including its role in training AI models without labelled data.

• Explore Generative Adversarial Networks (GANs) and how they are used in data synthesis and augmentation.

• Master Probabilistic Modelling and Bayesian approaches for uncertainty estimation in machine learning models.

• Understand the applications of probabilistic techniques in data analytics, decision-making, and predictive modelling.

Instruction Methods

• Lectures via Zoom using slides

• Practical exercises with Google Colaboratory

• Interactive polls and Q&A sessions

Assessment

• Participation in class, discussions and polls

Course Material

The Zoom codes to view the recordings are available in the Introduction Lecture.
Enrolled students who need access to course materials should contact me via email.

• Lecture 1 and 2: Introduction
  Zoom recording 1, Zoom recording 2

  • Colaboratory Notebooks:

    • Data Augmentation

    • Hyperparameter Tuning with Tensorboard

    • Exploring the MNIST dataset with a Convolutional Neural Network

    • From Blurry to Sharp: Super-Resolution for MNIST

    • Techniques for handling imbalanced datasets

    • Latent Space Extraction and Classification with XGBoost

    • Finding outliers in data through UMAP and HDBSCAN

    • Latent Space Visualization for a Regression Task

    • Interpreting Time Series Classification with SHAP

    • Cats vs. Dogs Classification Using the Pre-Trained CNN-based MobileNetV2

    • Exploring MNIST Latent Spaces: PCA, t-SNE, and UMAP on CNN Features

• Lecture 3: Generative Adversarial Networks, Transformers, Autoencoders
  Zoom recording 3, Zoom recording 4

  • Colaboratory Notebooks:

    • GAN application to the MNIST dataset using PyTorch

    • WGAN-based refinement of simulated data

    • Image-to-image translation with CycleGAN

    • Enhanced Time Series Classification with CNNs and Pre-trained Embeddings

    • Using the DINOv2 META AI transformer-based model for a galaxy classification task

    • Dynamic Masked Autoencoder with Linear Layers

    • Dynamic Masked Autoencoder with a CNN architecture

    • Transformer-based autoencoder with Consecutive Chunk Masking for Time Series Reconstruction

    • MNIST Autoencoder: Image Reconstruction & Latent Space Analysis

    • Denoising Autoencoder for MNIST Digit Restoration

• Lecture 4: Graph Neural Networks
  (still Zoom recording 4)

  • Colaboratory Notebooks:

    • Graph Classification using a GNN

    • Paper type prediction using a GNN

• Lecture 5: Probabilistic Learning, Bayesian Deep Learning, Variational Autoencoders
  Zoom recording 5

  • Colaboratory Notebooks:

    • Exploring Bayesian Regression for Uncertainty Quantification in Predictions

    • Variational Autoencoder on MNIST data

• Additional Material
  (still Zoom recording 5)

  • Colaboratory Notebooks:

    • Exploring High-Dimensional Representations with UMAP and Inverse Transformations