Training the Archive

Various ANN that already contain mathematical weights can be used via public libraries such as Keras or TensorFlow. The complex process of training the weights is based on the ImageNet image database, which is composed of up to 14 million images from the Internet. This has led to questionable or even biased categorisation.

Short for »International Image Interoperability Framework.« A standardised interface, e.g. for the inter-institutional exchange of image data and other digital objects.

Means the classification of various objects in a data set into different groups. The classification is carried out automatically on the basis of detected similarities, e.g. in an image corpus for groups such as »dogs« and »cats.«

Describes a disproportionate weight, e.g. in the training of AI, in favour of or against information contained in the data. Conversely, this can result in disadvantages or perpetuate unfair biases, which is particularly critical when the data is used as the basis for making decisions with (in)direct impact on daily life.

Short for »Application Programming Interface.« Stands for a programming interface that enables the connection of a piece of software to another program, e.g. for scraping data sets of museum collections.

Humanisation, i.e. human characteristics are attributed to the non-human. The machine is supposed to be intelligent like humans (or surpass them) and in the process obtain a circuitry that resembles the human brain.

Annotated, i.e. provided with a note. For example, knowledge in the form of metadata and tags can be added to specific images or digital objects in order to better classify or filter them.

A sequence of instructions for solving a problem. Algorithms follow defined individual steps, which are executed in their specified order (input, processing, output).