Natural language processing, robotics, video processing, stock market forecasting and other similar tasks require models that can deal with sequence data and understand temporal dependencies. Two major classes of models that have been designed to deal with sequence data are recurrent neural networks (RNNs/LSTMs) and transformer architectures. Designing and understanding these models is a very active and diverse area of research. Applications of these models are also widespread. The recent explosion of interest in topics such as language modelling and machine translation is based on advances in these models which includes GPT-3, DALL-E, etc.
In this course you’ll first understand the fundamentals of recurrent neural networks and transformers that led to these breakthroughs. Then we’ll go through and discuss both seminal and recent research papers on these topics to throw light on algorithms and challenges in this field.
The allocation of the limited seminar places is done via the Moodle course "Masters Seminar: Topics in deep learning for sequence processing" for the corresponding semester. Registration for the winter semester is possible until 11.10.2021. Please enter your degree program in the comment field. Places are allocated with priority to Master students of Applied Computer Science, but the course is open to students of other faculties who fulfill the content requirements.
- Course type
- Winter Term 2021/2022
moodle course available
every week on Wednesday from 14:00 to 16:00.
First appointment is on 13.10.2021
Last appointment is on 02.02.2022
We expect a solid level of mathematics as taught in the Applied Computer Science Bachelor‘s. Tools commonly used in machine learning are
- basic probability theory/statistics (expectations, variance, foundational distributions and densities, markov chains)
- linear algebra (matrices, vectors, eigenvalues/eigenvectors)
- calculus (functions, derivatives/gradients, simple integrals)
The course material is in English, the course language will be English.
Taking the course on Artificial Neural Networks either before or in parallel to this course is recommended (but not required).