My dissertation investigates the epistemology and aesthetics of interpretable machine learning, in particular with regard to feature visualization. Other research interests include the preservation of digital art and the interface of machine learning and performance art. My dissertation committee consist of Prof. George Legrady (Experimental Visualization Lab, MAT), Prof. Wolf Kittler (Comparative Literature), Prof. Alan Liu (English), and Prof. Marko Peljhan (Systemics Lab, MAT). I am also part of the WE1S humanities advocacy project at UCSB, and have been affiliated with the STAGE lab at the University of Chicago.

Curriculum Vitae / Google Scholar Profile

Dissertation abstract

Interpretable machine learning has recently attracted a significant amount of interest from research communities within and outside of computer science. From the technical perspective, machine learning systems become interpretable if they, either by design or with the help of external tools, provide human-understandable explanations for their decisions. For deep convolutional neural networks (CNNs), the iterative optimization of images has become the most widely used among these external tools. However, while the ambiguous nature of such visualizations has been noted in the computer science community from its inception, to date there has been no critical analysis of CNN visualizations in media studies, visual studies, science and technology studies, or related humanist disciplines. This is surprising in so far as the visualization of CNNs poses a number of epistemic and aesthetic problems that are closely related to similar problems in the humanities. Among them are the problem of perception (how is reality processed visually?), the problem of representation (how does an image represent reality?), and the problem of interpretation (what is the relation of form and meaning in an image?). The dissertation attempts to close this gap between technical reality and critical reflection by asking the question: What are images of neural networks images of? Specifically, considering the visualization of CNNs as a contemporary scientific image-making technique, the dissertation posits that the representational capacity of CNN visualizations, i.e. their capability to be images of something, depends significantly on additional visual information introduced into the machine learning system. This has problematic consequences for real-life applications of CNN visualizations, like the recent use of feature visualization for the reconstruction of images from human brain activity, which inherits a significant epistemic and aesthetic bias from its dependence on natural image priors.

Custom class activations for Inception V3

Peer-reviewed conference papers

   Fabian Offert, "Images of Image Machines. Visual Interpretability in Computer Vision for Art", in: Proceedings of the European Conference on Computer Vision, Springer, 2018

Book chapters

   Fabian Offert, "Hands on Circuits. Preserving the Semantic Surplus of Circuit-Level Functionality with Programmable Logic Devices", in: Hands on Media History, ed. by John Ellis, Nick Hall, Routledge, forthcoming

   Fabian Offert, "Beyond the Scenes. Sasha Waltz's Objects and Installations Between Theater and the Visual Arts", in: Sasha Waltz. Installations, Objects, Performances, ed. by Peter Weibel et. al., Ostfildern: Hatje-Cantz, 2015   

Exhibition catalogs

   Open Sources, ed. by Fabian Offert and Juan Manuel Escalante, Santa Barbara: University of California, Santa Barbara, 2015   

   Jonas Mekas: 365 Day Project, ed. by Fabian Offert, Karlsruhe: ZKM | Center for Art and Media, 2014

Research Blog Posts

   Fabian Offert, Teddy Roland, Devin Cornell, "Word Embeddings for Restricted Access Corpora", 2018


   Fabian Offert, "Notes on the Aesthetics and Epistemology of Interpretable Machine Learning", University of California, Santa Barbara, October 2018

   Fabian Offert, "Notes on the Aesthetics of Artificial Intelligence", Exploring Edges: International Colloquium on the Digital Humanities, Architecture, Artistic Research, and Criticial Technical Practice, EPFL Lausanne, July 2017

   Fabian Offert, "Preservation as Translation. The Case for Programmable Logic Devices as a Strategy for Circuit-Level Authenticity", International Workshop on the Material Authenticity of the Ephemeral, Deutsches Museum Munich, October 2017

   Fabian Offert, "Exhibiting Computing Machines. Computational Agency as a Speculative Principle for Exhibition Design", SPT2017: The Grammar of Things, TU Darmstadt, June 2017

   Fabian Offert, "[machine here]. Non-representation as a Rhetorical and Mathematical Strategy in the Construction of Alan Turing's Imaginary Machine", [image here] conference Harvard University, 2016

   Fabian Offert. "The Conservator's Task. The Case for Programmable Logic Devices as a New Tool for the Conservation of Digital Art". Hands on History conference, University of London, 2016

   Fabian Offert, "Re-enactment as Pre-enactment. The State of Emergency and the Theatrical Contract in Artistic Practices of Reenactment", Radical Ephemeralities Conference, University of California, Santa Barbara, 2015


   Interview on SWR2 (in German), 2013   

   Fabian Offert. "Information, Konzept, Berechenbarkeit. Zum Computer als Medium der Kunst" (diploma thesis, in German), Gießen: Justus-Liebig-Universität, 2012