How smart is a smart home?

Smart home technology is used in a growing number of households. But how smart is a smart home really and what potential does Machine Learning hold in developing truly intelligent household technology? As part of the Science Code Slam, ML2R scientists analyzed a previously unexplored data set, encompassing over 50 million individual events from three years. They succeeded in detecting anomalies in the data set and identified potentials for intelligent applications.

AI for Jokes: Creative Machine Learning for the generation of Jokes

To understand a joke, the human mind must think actively and creatively to retrieve, associate, and construct knowledge. Can Artificial Intelligence also generate a joke? ML2R researchers addressed this question as part of the Science Code Slam. They implemented a novel approach to Machine Learning that combines neural and symbolic approaches and has been developed within ML2R. By establishing an interactive website, the team illustrated the complex web of knowledge and emotions that come into play in the context of jokes.

Down to the (Bayesian) basics

Autoencoders are artificial neural networks that learn representations of data which make the data more easily usable and understandable. Using artificially created datasets, researchers investigated how to generate new, statistically similar data by using trained Autoencoders in an iterative fashion. They first randomly chose a starting representation and then decoded and encoded it repetitively. After the last decoding step, the researchers had obtained new artificial data, statistically similar to the training datasets.

Explain this to me! Interactive Website for Neural Network explanations

(Deep) Neural networks have become the go-to solution for practitioners when it comes to implementing Machine Learning approaches on their data. Especially in safety-critical areas, for example autonomous driving, explanations for the predictions of a model are needed. ML2R researchers programmed an interactive website that allows users to selectively modify points within a data set. In real time, users receive visualized explanations of the resulting changes in model behavior.

Hack the traffic: The dangers of adversarial attacks on ML technologies

Especially in high-risk application areas, IT security is of vital importance for ML technologies. Hostile attacks on prediction models, for example in road traffic, could cause serious damage. This was demonstrated by ML2R scientists using a data set on traffic signs. They also explored different defense mechanisms against said attacks.

Communication is Key: How to write a good Blog Post

Making Machine Learning and Artificial Intelligence technologies and research results accessible to companies and society: With this goal in mind, ML2R will launch its German-language blog on Machine Learning in 2021. In a Science Code Slam session, participants wrote blog posts together and addressed the requirements of successful science communication.

Untangling the hidden potential of Quantum ML

At the Competence Center ML2R, scientists are researching the potential of quantum technology for Machine Learning. Quantum computers can significantly accelerate the execution of ML processes. A working group of the Science Code Slam designed a hybrid model combining aspects of classical ML and novel quantum technology. The model was then trained on a real-world data set.