https://github.com/defgsus/neurost

Neuronal ShaderToy Framework
https://github.com/defgsus/neurost

Last synced: 28 days ago
JSON representation

Neuronal ShaderToy Framework

• Host: GitHub
• URL: https://github.com/defgsus/neurost
• Owner: defgsus
• License: other
• Created: 2016-02-09T13:52:15.000Z (almost 9 years ago)
• Default Branch: master
• Last Pushed: 2016-02-10T14:42:37.000Z (almost 9 years ago)
• Last Synced: 2024-10-10T15:49:50.671Z (about 1 month ago)
• Language: GLSL
• Size: 21.5 KB
• Stars: 4
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Neuronal ShaderToy Framework

Experiments with multi-layer neural networks on http://shadertoy.com

### status

- classic multilayer perceptron with states, weights, and errors in only two render buffers 

- basically working for up to 4 layers (2 hidden)

- max number of layers is fixed in code but can be expanded. The lack of const arrays in webgl requires a lot of workaround. Also limited by screen resolution

- time inefficient: 4 layers need 8 frames for one training case: only 7.5 updates per second. GPUs could do *much* better

- hard to analyze performance because of slowness..

- can save two frames by not copying input state and pre-calcing error, which would make code less coherent for layers