Artificial neural networks (ANNs) denote a popular class of models used within machine learning. An ANN typically consists of multiple layers of simple processing units, so-called artificial neurons. Most current ANNs involve multiple layers of these processing units, hence the term deep learning is sometimes applied to describe them. Historically, they emerged from a neurophysiological inspiration to express the processing of mammal neurons in mathematical terms (cf. McCulloch and Pitts 1943). There exists a plethora of different approaches to the design of ANNs, some variations include the number of artificial neurons in a layer, the activation function applied, or the connection of artificial neurons between layers. From these variations have emerged several classes of ANN architectures, such as Multi-Layered Perceptrons (MLPs), Generative Adversial Networks (GANs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), or more recently Transformers. It is conspicuous that many papers, which discuss a particular ANN architecture, represent them in some diagrammatic form. This diagrammatic representation, however, does not follow any unified structure. This results in two challenges: First, ANNs are not visually comparable through an analysis of their diagrammatic representations. Second, the depicted diagrams of ANNs might lack relevant information, overseen by the original researchers. In short: It appears that the depiction of ANNs lack a conceptual modeling language.

The present thesis should adress this gap. Therefore, it is relevant to expound on the foundations and variations of ANNs as well as to explore the fundamentals of conceptual modeling languages. Based on an analysis of the design, evaluation, and application of ANNs, requirements for a corresponding modeling method should be derived. Thereupon, these insights should be used to specify a conceptual modeling method for ANNs.

Literature:

• Aggarwal CC (2018) Neural Networks and Deep Learning: A Textbook. Springer International Publishing: Cham
• Du K-L, Swamy MNS (2014) Neural Networks and Statistical Learning. Springer-Verlag: London
• Frank U (2013) Domain-Specific Modeling Languages – Requirements Analysis and Design Guidelines. In: Reinhartz-Berger I, Sturm A, Clark T, Wand Y, Cohen S, Bettin J (eds.) Domain Engineering: Product Lines, Conceptual Models, and Languages. Springer: Cham, pp. 133-157
• Kelleher JD (2019) Deep Learning. The MIT Press: Cambridge, MA, London
• McCulloch WS, Pitts W (1943) A Logical Calculus of the Ideas Immanent in Nervous Activity. Bulletin of Mathematical Biophysics 5:115-133

If language shapes our reality, changing our language might lead to a different, potentially preferable reality. This thought is echoed throughout a variety of philosophical schools and can, in different variations, with different assumptions, and with different implications, be found in the writings of Ludwig Wittgenstein, Richard Rorty, Friedrich Nietzsche, Immanuel Kant, or Humberto Maturana. Recently, the discussion has received more widespread attention. Motivated in part from feminist philosophy of the 1990s, philosophers have combined their research efforts towards the improvement of language under the moniker of conceptual engineering and conceptual ethics. The amelioration of concepts and language is faced with several theoretical and practical challenges. What makes a concept “better” than another? How could a new concept be adopted by respective language users?

Information systems development is essentially concerned with language development (clarification and sources per request). Broadly, this poses the question if information systems can support conceptual engineering and, if so, in what regards. Machine learning (ML) might be a fruitful first step to guide this analysis. Contemporary ML approaches are inductive (cf. Rescher 1980): they generate potentially novel generalizations based on a set of observations. Researchers like Rees (2022) therefore suggest that they might guide the development of novel concepts.

This master’s thesis should explore the capabilities of ML to support conceptual engineering. You should identify potential tasks of conceptual engineering and what requirements they face. Then you should investigate how different ML approaches (we can disucss which in our first meetings) can serve to address these requirements.

Literature:

• Burgess A, Cappelen H, Plunkett D (eds) (2020) Conceptual Engineering and Conceptual Ethics. Oxford University Press: Oxford
• Butlin P (2021) Sharing Our Concepts with Machines. Erkenntnis
• Cappelen H, Dever J (2019) Bad Language. Oxford University Press: Oxford
• Haslanger S (2012) Resisting Reality: Social Construction and Social Critique. Oxford University Press: Oxford
• Medin DL, Smith EE (1984) Concepts and Concept Formation. Annual Review of Psychology 35(35):113–138
• Montemayor C (2021) Language and Intelligence. Minds and Machines 31:471–486
• Ontañón S, Dellunde P, Godo L, Plaza E (2012) A Defeasible Reasoning Model of Inductive Concept Learning from Examples and Communication. Artificial Intelligence 193:129–148
• Rees T (2022) Non-Human Words: On GPT-3 as a Philosophical Library. Daedalus 151(2):168–182
• Rescher N (1980) Induction: An Essay on the Justification of Inductive Reasoning. Basil Blackwell: Oxford