Deep learning has proven as one of the most powerful machine learning strategies having emerged so far. A prominent example is image analysis, where performance rates in classification exceed human performance rates. Since its early successes, deep learning has evolved into a highly developed machine learning strategy, which encompasses a variety of different techniques, each of which addresses particular, often challenging issues.
Deep learning refers to classification of input data by means of a neural network that consists of an elevated number of 'network layers'. The exact shape of these layers, and the composition of layers of different shapes, makes the architecture of a deep neural network. In this crash course, we will, after having discussed the basics and foundations of machine learning in general, and deep neural networks in particular, move on by studying different architectural elements and highlight their strengths, and also their possible weaknesses. Along with studying (also somewhat more advanced) network architectures, we will introduce popular deep learning software frameworks, as practical support for relevant exercises. Last but not least, we will highlight applications in biology, where the recent surge of data intense experimental technologies has pointed out the great potential benefits of deep learning.
Contents. Machine Learning, Supervised Learning, Deep Neural Networks, Regularization, Optimization, Backpropagation, TensorFlow, Pooling, Autoencoders, Convolutional Networks, Recurrent/Recursive Networks, ResNet, Inception (GoogleNet), Generative Adversarial Networks, Applications in Biology
Material.
- I. Goodfellow, Y. Bengio, A. Courville: Deep Learning. MIT Press, 2016. Web resource: <www.deeplearningbook.org>
- 10 Advanced Deep Learning Architectures Data Scientists Should Know: <https://www.analyticsvidhya.com/blog/2017/08/10-advanced-deep-learning-architectures-data-scientists/>
Prerequisites. Familiarity with elementary concepts from linear algebra, probability and numerics are helpful (but not mandatory). Prior knowledge in programming Python is desirable.
Format. Lectures will alternate between introductions to concepts and theory of complex systems and more practical elements referring to implementation of deep learning in practice. There will be regular, weekly exercises, an extended programming project, and a final exam.
Learning goals with assessment weighting:
- read and demonstrate (in class discussions) understanding of literature and lecture contents (20%)
- understand and be able to apply deep learning in exercises and programming project (40%)
- demonstrate understanding of theoretical concepts in final exam (40%)
Evaluation matrix:
| |
in class discussion
20% |
Exercises/ project report 40% |
final exam 40% |
| is able to read and demonstrate (in class discussions) understanding of literature and lecture contents |
x |
|
|
| understands and is able to apply deep learning in exercises and programming project |
|
x |
|
| is able to demonstrate understanding of theoretical concepts in final exam |
|
|
x |
|
 |
|
|