Neural networks-long believed dead-are back. Advances in training deep neural network architectures have had a tremendous impact on Computer Science and Machine Learning for more than a decade now. Today, they are the foundation of many modern data science applications in academia and industry.  Their capacity to handle unstructured information such as text, audio, or image data makes them particularly attractive for researchers who want to solve research problems that involve data beyond 'just' spreadsheets. This hands-on course is designed to demystify the complexities of deep learning, providing you with a solid foundation in its fundamental concepts and practical applications.

 

Deep learning, a subset of machine learning, involves training neural networks to learn and extract variables from complex data. Deep learning can be utilized in social science research in many ways and is the foundation of many modern tools, e.g. for analyzing text. For example, it is possible to employ deep learning for sentiment analysis, topic modelling, entity recognition, and language translation, all of which are valuable in understanding public opinion, political discourse, and social media behavior. Moreover, deep learning can also be the foundation of image and video analysis, social network analysis, time-series analysis, predictive modelling, or enhanced data pre-processing such as multiple imputation.

 

Deep learning is a little different from other, more classical machine learning models approaches. While classical machine learning models have fewer parameters and a shallower structure, deep neural networks often consist of many layers with a large number of parameters. Their makeup allows them to learn highly complex and hierarchical features from the data. Deep learning can thus automatically retrieve features from the raw data through the network's layers, reducing the need for extensive manual feature engineering.

 

This workshop will provide an introduction to deep neural networks. Starting from flat architectures, we will understand how neural networks use backward propagation of prediction errors to optimize their performance. Adding more layers to the initial set-up, we will then go on and build increasingly “deep” architectures. We implement fully connected deep learning architectures and learn about classical pitfalls and how to tackle them.

 

The workshop will also give up the full connectedness between nodes and take a first look at models that aggregate information to more abstract data representations and recurrent data. We will guide you through the fundamental principles of deep learning and its application in various domains, focusing specifically natural language processing (NLP) with a brief excursion to computer vision at the end of the workshop. The course is structured into an introduction to the basics of deep learning and two main sections: