Alignment work is a crucial part of developing and maintaining AI systems. In the broadest sense, aligning an AI system describes the correction of undesired behavior of the AI by creating rule sets and extensions within an existing AI data set.
Prototype I: Deep dive the AI at convolutional level
The visualization of an artificial neural network and the associated interaction with the network by means of virtual reality enables direct access to these often very complex networks, which we can often only perceive as black boxes.
This first prototype shows three convolutional layers that are already connected to each other. Touching a node visualizes its connections through all layers. With this approach, the „deepest“ level of an AI system is made visually accessible, as only the patterns of the data are shown structurally, but no more readable, more abstract form has yet been chosen.