Thinking - Robert Hamill

Neural Network Fundamentals (cont)

external reality. This is important because it limits the multiplicity of reality so that your computing intelligence is able to categorize and form hypotheses about similarly categorized reality.
   Because each neural node performs its task when its inputs drive it, all the neural nodes can act simultaneously. This parallelism is important for speed and it is achieved without an overall programming language.
   Image recognition allows an excellent comparison between neural networks and von Neumann (single process) computers. Physiologists have determined that human neurons transmit their messages in ten milliseconds. Since we can recognize a face in a second, it is estimated that the longest programmatic construct is less than a hundred steps.
   Since the human eye condenses 125 million photoreceptors to 1 million ganglion cells that go to the brain's occipital lobe, it's obvious that the immense parallelism of the neural net is what allows our rapid image recognition.
   For artificial intelligence projects, these concepts

are selected for the domain of the problem set. For example, for an assembly line robot, shapes of the parts being assembled and relative locations of the parts. For an audacious project, a robot in the Isaac Asimov mode, the concept table would require hundreds of words, thousands of images, and a multitude of relationships between them.
   It's important to consider how the members of the concept table are added, changed, or deleted. In an advanced robot, not all concepts would be built in. You'd want it to learn. Learning in neural networks will be the subject of another essay.
   In short, learning is training it to recognize relationships that we want established. Learning involves establishing linkages in the neural net between inputs and desired outputs, by repeatedly presenting the neural net with the input and slowly adjusting the internal weightings between the neurons to come closer to the desired outputs.
   In neural networks learning is complicated by the fact that the same neurons are used for multiple recognitions and linkages.
   Since this was theoretical, I'm skipping the difficult engineering aspects.