Seminar paper by Florian Rottach 2018 (FZI&KIT) – Data Science and Real-Time Big Data Analytics
In the domain of machine learning, new methods such as deep learning have made enormous progress in recent years, which is particularly evident in the classification of data. The models used are becoming more accurate through adaptations and categorize data at hand with decreasing error.
A well-known challenge that can arise in supervised training is the lack of classified data. As a result, the training set is not extensive enough to provide the learning engine with all the necessary parameters. In the course of this, learning procedures have been developed which allow the generation of data that can be used as training data, among other things
Instead of collecting, extracting and transforming arbitrary data, the idea is to specifically produce the desired instances. For this purpose generative models are used, which can generate similar, but completely new instances from given data. In the past, these approaches have not been very successful, which is mainly due to the structure of previous methods. In this paper, Generative Adversarial Networks (GANs) are presented, which yield significantly better results compared to previous methods. In particular, the effective generation of new data drives the implementation of machine learning creativity.
This creativity in machine learning can be a fundamental factor in implementing artificial intelligence similar to that of humans. This offers the possibility of transferring a wide variety of design tasks to these models. For example, one could imagine independently writing a book based on similar works.
Until now, humans have been distinguished from intelligent machines primarily by the fact that they possess creativity and can develop completely new ideas and approaches. However, through generative models, AI systems are increasingly able to adapt this ability. This specifically raises the question of what consequence follows from the fact that artificially created data can no longer be distinguished from the real thing.
Excerpt from the seminar paper by Florian Ottach
The question that arises for us is the following: Can we use AI to automatically create an artificial document that comes close to a real safety data sheet in terms of content and appearance? We also know that the generated SDS will not stand up to plausibility checks, but that is a topic for the future. We are interested in finding out the potentials of GANs for the topic of SDSs, so that we can use them in a targeted way. Well, Florian’s work is a great basis, because he showed that this approach worked great for generating handwritten character sets.
We will continue to pursue this topic in the future and are looking forward to the insights we will gain in this context.