Dmytro Panchenko
13 Oct 2018
17:00 - 18:40

Tuning CNN: tips & tricks

Nowadays implementation of neural networks with high-level frameworks is pretty easy, but tuning a deep network to achieve top accuracy might be tricky. In this workshop we will discuss tips and tricks which can be applied to training deep CNN. As an example we will take a usual classification task to demonstrate step-by-step improvement of model’s quality. During the first part of the workshop we will consider basic things such as:
1. Transfer learning
2. Learning curves interpretation
3. Learning rate management
4. Augmentations

In the second part we will discuss in theory and try in practice more advanced tricks such as:
1. Working with imbalanced data
2. Test-time augmentation
3. Semi-supervised approach (pseudolabeling)

Minimal requirements to the listener: general knowledge of neural networks, basic knowledge of Python.
Dataset and pre-trained models: