Image influx

Visual Artificial Intelligence systems, such as those which generate pictures of dogs or analyse their contents, rely on large collections of images known as training datasets. A dataset used to train an AI system is incredibly vast and, as such, hard to introspect: each image in it is collected, classified and labelled, and different datasets may do this in different ways. The decisions over collection, classification and labelling of images can greatly change how the AIs trained on them operate in the world.

As part of this week-long workshop, we will play games and carry out exercises to consider how different decisions lead to the production of different training datasets, exploring processes for image collection, classification and labelling. Participants will be asked to create their own image dataset accompanied by a datasheet clarifying and justifying its creation process and possible uses. The dataset and datasheet will be presented online at the end of the workshop.

Topics we will explore
—Labelling: images do not describe themselves and need to have their contents labelled. How do humans tell computers which words will relate to a given image? How do they decide what an image means?
—Classification: images in datasets need to be classified, and this may rely on existing classification systems inherited from other domains. How much does the interpretation of an image depend on the way the dataset is classified?
—Recursion: images for datasets are often automatically collected in large quantities from the web. At the same time, AI systems lead to the mass generation of images that are themselves hosted on the web. How may these AI-generated images impact future datasets?
—Simulation: due to the large amount of images required, datasets rely on images rendered or exported from real-time 3D simulations running on video game engines, rather than being collected from the web. What are the implications of simulating the real world to train AIs to operate in it?