Description
Since their earliest conception in the 1940s, artificial neural networks have been alternatively regarded as extremely promising machine learning models, capable of learning anything, and as glorified linear combinations, unable to achieve relevant results in practice.
However, along the last decade, the availability of general-purpose GPU architectures and large quantities of data has enabled the rise of deep neural networks, which have attained state-of-the-art performance in many applications, from image classification to text translation. This has given rise to a whole new field of research, ranging from generative models to adversarial attacks (and defenses against them).
This course is intended as a first contact with artificial neural networks, followed by an overview of the different architectures that are currently available.
This course is part of a larger course series in Data Analysis consisting of 19 individual modules. Find more information and enroll for this module via www.ipvw-ices.ugent.be
Program
- Introduction to neurons and neural networks
- Training with backpropagation
- Challenges and solutions to train deep neural networks
- Convolutional networks
- Adversarial examples
- Generative models: Autoregressive models, Autoencoders, Variational autoencoders (VAE), Generative adversarial networks (GAN)
- Transformers and BERT
- Recurrent neural networks
The practical sessions use the Python library TensorFlow to implement some of the models discussed in the course, with particular emphasis on how to adapt the networks to the characteristics of a specific problem.
