© Jake Elwes
- Artist/s
- Jake Elwes
- Title
- Closed Loop
- Year
- 2017
- Edition / Serial number
- 5 + 2 AP
- Copy Number
- 3
- Category
- Video
- Format
- Video Installation
- Material / Technique
- computer-generated video installation; 2 projectors, 2 videos, color, without sound; production video: computer: Intel Core i5, Nvidia GTX 1080ti, 16GB RAM, operating system: Ubuntu 16.04,
programs: individual software, DenseCap, generative neural networks, programming languages: Python, Tensorflow, Caffe, Torc
- Dimensions / Duration
- installation dimensions variable
- Collection
- ZKM | Center for Art and Media Karlsruhe
- Description
- The video »Closed Loop« documents the conversation between two artificial intelligences. One artificial neural network communicates through written words, the second artificial neural network through images.
The language-based network describes the image that has just been generated by the other network. The image-generating network in turn now tries to generate an image that matches the image description. The two systems alternately display the content of the other - as best they can.
Occasionally there are aberrant interpretations, one could also say misunderstandings, so that the conversation abruptly develops in a completely new direction. Are the possibilities of mutual understanding of two AI systems very different from those of two communicating people?
The systems were trained with the help of large data sets. The speech-based algorithm was prepared to identify images using 4.1 million images, while the image-based algorithm was trained to respond to words using 14.2 million photographs from the ImageNet database. After completing the training processes, the two systems were able to produce images and words completely autonomously.
Author
Margit