Self-organizing neural network architectures for computing visual depth from motion parallax

The analysis describes some of the issues involved in constructing a self-organizing neural network that can learn to perform a high-level vision task, depth perception from motion parallax, without guidance from an external teacher. An examination is made of how motion parallax conveys depth information. A network structure is presented for detecting and representing depth from motion parallax. How a depth-sensitive network can self-organize is also examined.<>