Images of Image Machines. Theory and Practice of Interpretable Machine Learning for the Digitial Humanities

While text is still the most important research topic in the digital humanities, over the past ten years images have started to gradually appear on the radar of computational humanists. Recent developments in digital art history in particular have shown that the importance of images for DH research goes beyond ensuring their accessibility through databases and interfaces. In fact, images are where digital humanities and artificial intelligence meet. Most importantly, the automated classification of images on the one hand, and the automated production of images on the other raise a fundamental question at the interface of computer science and the humanities: how is reality represented in machine learning systems? The field of interpretable machine learning is concerned with opening the black box and answering this question.

This two-week workshop will serve as an introduction to the theory and practice of interpretable machine learning. The first week will introduce participants to the field by means of reading, discussing, and replicating foundational results in interpretable machine learning, with a particular focus on the fairness, accountability, and transparency (FAT) of machine learning systems (see for instance: http://gendershades.org/). The second week will be dedicated to hands-on experimentation with image datasets in PyTorch (https://pytorch.org/), a popular machine learning framework. While the first week has no prerequisites, the second week requires basic programming skills, preferably in Python. Topics include but are not limited to:

First Week

• The current state of artificial intelligence
• Interpretable machine learning and artificial intelligence
• What is it like to be a machine - phenomenology of machine learning
• Images vs. text: "continous" vs. "discrete data"
• ImageNet: an awful dataset that is used everywhere (and its history)
• How to make a racist AI without really trying

Second week

• Introduction to PyTorch
• Introduction to art historical image datasets
• Convolutional neural networks for image classification
• Generative adversarial networks for image production
• Feature visualization and Deep Dream: visualizing learned features of neural networks
• Experiments with CNNs and GANs: generating millions of images for interpretation