Deep Learning

Presentation

Deep learning is a technology that is booming, thanks in particular to the use of GPUs (Graphical Processing Units), the availability of large amounts of data and the understanding of theoretical elements that make it possible to better define neural network architectures that are more easily trainable. In this course, students will be introduced to the basics of neural networks and also to the different architectural elements that make it possible to design a neural network according to the prediction problem considered. The course is divided into modules in which questions of optimization algorithms, their initialization, regularization techniques, fully connected architectures, convolutional networks, recurrent networks, introspection techniques are addressed. Practical works on GPUs are associated with the courses.

Learning outcomes

• Being able to implement and deploy a deep learning algorithm
• Being able to choose the right architecture that suits a particular machine learning problem
• Being able to diagnose the training of a neural network (what is it learning ? how is it learning ? is it learning ? will it be able to generalize ?)

Syllabus

• Historical introduction to neural networks, linear classifer/regressor (1.5 HPE)
• Computational graph and gradient descent, Fully connected networks, RBFs, Auto-encoders, Optimization algorithms, initialization, regularisation (3 HPE)
• Convolutional neural networks : architectures (1.5 HPE)
• Convolutional neural networks: classification, object detection, semantic segmentation (1.5 HPE)
• Recurrent neural networks: architectures and training (1.5 HPE)
• Recurent neural networks: applications (1.5 HPE)
• Introduction to generative and probabilistic networks (RBM, Deep Belief Networks) (1.5 HPE)

The praticals will be on:

• Introduction to pytorch on classification with linear predictors, fully connected networkjs and convolutional networks (3 HPE)
• Convolutional neural networks for object detection and semantic segmentation (3 HPE)
• Recurrent neural networks applied to sequence to sequence translation (3 HPE)
• Generative networks (3 HPE)