fr
en
Understand the theoretical foundations and algorithms of deep learning
Deep learning becomes usable when you understand the choices that influence its results. Connect architectures, data and training to analyse model behaviour. Develop a framework for designing experiments and assessing their performance rigorously.
- Duration
- 2 days 14 hours
- Code
- IA040FR Code
Presentation
Deep learning is often used as a black box through high-level libraries. Yet innovation and fine-tuning models require an understanding of what happens under the hood. This 2-day course provides an immersion in the mathematics and algorithms that underpin deep learning.
The programme moves beyond convenient abstractions to return to the fundamentals: matrix calculations, gradient descent and backpropagation. You will learn to implement a neural network from scratch in Python, without a framework, to understand the inner mechanics of learning.
You will then progress to modern architectures, including CNNs and RNNs, and optimisation strategies such as Adam and RMSProp. Combining theoretical rigour with practical implementation, this approach equips you to diagnose, adjust and design high-performing models with genuine technical understanding.
Objectives
By the end of this course, you will be able to:
- explain the mathematical foundations of deep learning, including linear algebra, derivatives and gradients;
- master forward propagation and backpropagation mechanisms;
- implement optimisation algorithms such as SGD and Adam manually and understand their hyperparameters;
- distinguish and apply neural network architectures, including MLP, CNN and RNN, according to the data;
- evaluate model performance and address overfitting or underfitting.
Program
Day 1: Mastering theoretical foundations and basic algorithms
- Mathematical foundations: vectors, matrices, dot products and activation functions, including sigmoid and ReLU.
- Learning mechanics: cost functions, forward propagation and gradient calculation through backpropagation.
- Optimisation algorithms: gradient descent, momentum, RMSProp and Adam.
Hands-on exercises
- Manually calculate a propagation step and implement a simple neural network in plain Python, without a framework.
Day 2: Exploring advanced architectures and applying best practices
- Specialised architectures: convolutional networks (CNNs) for images and recurrent networks (RNNs/LSTMs) for sequences.
- Training and evaluation: managing overfitting, cross-validation and performance metrics.
- Best practices: weight initialisation, normalisation and an introduction to transformers.
Hands-on exercises
- Implement a CNN for image classification with Keras or PyTorch and tune hyperparameters to optimise performance.
Audience
This course is intended for those seeking to open the deep learning black box, including:
- data scientists and data engineers who want to move beyond superficial library use and build expertise;
- AI researchers and lecturers who need a firm grasp of mathematical foundations for their work;
- experienced developers seeking to specialise in implementing complex algorithms;
- artificial intelligence students seeking to consolidate theoretical foundations through practice.
Prerequisites
The following prerequisites apply:
- Technical skills: basic Python knowledge is essential.
- Theoretical foundation: knowledge of machine learning and statistics, and familiarity with the neural network concept, are desirable.
Teaching and assessment methods
- Initial skills assessment
- Training materials provided to participants
- Continuous assessment throughout the course
- End-of-course feedback questionnaire
- Combination of theory and practical application
- Attendance records
- Post-course follow-up evaluation
- Practical exercises
Course highlights
- From-scratch approach: code algorithms manually in plain Python for an in-depth understanding of their internal mechanisms.
- Scientific rigour: approach deep learning through its mathematical foundations, developing lasting skills beyond technology trends.
- Architectural perspective: understand not only how architectures work, but why CNNs and RNNs suit particular problems.
- Intensive practice: immediately validate each theoretical concept through implementation or a practical calculation.
Dates and sessions
Choose the date and delivery format that suit you.
No upcoming sessions are currently available.
Any brand names and logos mentioned in this course description, such as Python, Keras and PyTorch, belong to their respective owners. Their mention for educational purposes does not constitute an endorsement or partnership.