Design:Retina based neural learner

Raw pixel data is processed with convolution by a edge kernel. The result is then fed into a unsupervised prototype based leaner with a microfovea structure. The output of that is fed into another supervised prototype based learner to learn the basic category. The category is determined with by the label from a unsupervised prototype learner. The microfovea is inspired by the structure of a real fovea in human vision.

===> UL prototype learner with microfovea ===> SL prototpe learner
===> UL prototype learner -------------category---^

dev stage 1 (IMPLEMENTED)

• model has retina with a focal point
• retina resolution is on multiple levels which get more and more blury (there is evidence for that in the brain from neuroscience, TODO< search papers >
• prototype learning with some SDR'ish representation (?)
  • implementation the prototypes of the 2nd layer are implemented with NAL

dev stage 2

• implementation of a workspace to keep track of the scene
• control (as in a CPU) of memory system may be learned