Convolutional neural networks and image processing

Popis kurzu

Our workshop is for people who are looking for hands on experience with deep neural networks
for image processing, but they didn’t have any real opportunity to do so yet. Through
experiments, we will explore how and why such models work, what are the intuitions behind its’
functionality, and gradually, through simple examples, we’ll come to the models that are
commonly used in industry. We will focus on possible use cases for neural net’s internal semantic
image representation and how to visualize neural net behavior in the most effective way.

Our workshop is for people who are looking for hands on experience with deep neural networks
for image processing, but they didn’t have any real opportunity to do so yet. Through
experiments, we will explore how and why such models work, what are the intuitions behind its’
functionality, and gradually, through simple examples, we’ll come to the models that are
commonly used in industry. We will focus on possible use cases for neural net’s internal semantic
image representation and how to visualize neural net behavior in the most effective way.

Obsah kurzu

• Back to the history


• What the convolution is and why it works


• TensorFlow (designing a simple convolutional neural network)


• Practical classification task with the Fashion MNIST data set.


• Experiments with the MSCOCO and ResNet data sets


• Visualisations using TensorBoards


• Image classification


• How to deal with noisy data?

Předpoklady

• Základní znalost programování v Pythonu


• Středoškolské znalosti lineární algebry, matematické analýzy a teorie pravděpodobnosti. Bude předpokládáno základní porozumění pojmům jako vektor, matice, vektorový prostor, pravděpodobnost, podmíněná pravděpodobnost, nezávislost náhodných jevů a znalost násobení matic a derivace funkcí.


• Vlastní laptop s předinstalovaným Dockerem.


• Znalosti strojového učení na úrovni kurzu Úvod do strojového učení.

Studijní materiály