Hi BF33
Something has been bugging me about your theory, I think I have finally figured out what you are trying to describe.
Imagine a long horizontal growing memory bar/stick, divided into two rows horizontally.
First, Guess happens - actions for all the motors are generated and saved into the first box in the bottom row.
Then the actions are sent to motors.
Then senses go through rewards and are saved into the first top row box of the memory bar, above just tried actions. The senses were labelled and ranked by if they matched +/- rewards and how much they matched the reward.
The reason no one understands your theory is the way you are describing the system, your terminology is just wrong, and very, very miss leading. You also have a poor understanding of neural anatomy, programming and past/ current AI research.
Memory Front – I think your describing the state of mind in the moment or the current frame of consciousness.
Match box – This is a parallel connectionist neural net. That takes the individual sensory inputs (pixels from retina, etc) and recognises the individual pattern frames as a set of synaptic weights… in parallel. So each vision pixel, touch area value, etc is fed into the net individually but at the same time… in parallel. The best resulting pattern match is then reflected/ moved to the ‘memory front’. The ‘strength’ of the recognised pattern is then merged with new sensory data and sent back to the neural net for re-processing.
+/- Rewards – This is the altering of the weights on the synapse inputs to the above ‘match box’. You’re describing the tuning of the parallel net to recognise facets/ patterns of the incoming sensory streams.
Guess or random – Your describing the initial weights of synapse in the ‘match box’ parallel neural net. If no exact pattern is returned then your system injects random synaptic values which generates a random action, which may or may not be recognized in the next frame.
So for simple OCR it would look like this...
![](https://aidreams.co.uk/forum/proxy.php?request=http%3A%2F%2Fi.imgur.com%2F8ZEOGZF.jpg&hash=27d4db9f7db526b78a3c6e183c44a458d3157007)
If I'm right, what you are describing is a standard parallel connectionist recurrent neural model. Many companies are already working on this type of system and have been for many years, and is in fact the basic functionality of the current ‘Deep Learning’ craze.
https://en.wikipedia.org/wiki/Recurrent_neural_network