Possible AI configuration
Jun. 3rd, 2020 04:00 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
zeeth_kyrahJust some notes I'm sharing in case someone wants to try it out.
Neuron: Simple inversion reflex analog comparator, 0.2s memory per comparison segment.
Layer in a group of three layers (one grain): 1, 3, 12, cross-connected in a radial pattern. Pass-through buffer to center joining incoming links, pass-through buffer out of center joining outgoing links.
Allow each grain to read and compare specific data segments, filtered on the buffer in; no match, no pass.
Start with 1 grain minimum. That grain is MASTER and will never die. Add grains to input feeds where incoming signals "white out" active grains with overflow, remove grains from zones where signals live below reaction threshold; training is ongoing, with a slow build and slower decay - all grains can also change filter/mask. Remember you have a database backing up this information, use it wisely, this is simply an analysis unit for the database contents.
Assign types to grains: sensor, motor, thinker, memory. Technically memory are sensor neurons with internal feed and feedback. This assignment can be done randomly on a bias - every zone should have at least one of each, MASTER is always a thinker first, and is connected to a simple clock signal; note that this only changes external connections. Zones should each have at least one connection to another zone nearby (sourcing from a root provides this initially), and one zone a bit "further" away in the grouping.
Translate information incoming to this artificial brain into analog or analog-like signals, typically through a vocoder.
This should form a minimal, reactive artifical neuron brain capable of self-learning AI response.
Neuron: Simple inversion reflex analog comparator, 0.2s memory per comparison segment.
Layer in a group of three layers (one grain): 1, 3, 12, cross-connected in a radial pattern. Pass-through buffer to center joining incoming links, pass-through buffer out of center joining outgoing links.
Allow each grain to read and compare specific data segments, filtered on the buffer in; no match, no pass.
Start with 1 grain minimum. That grain is MASTER and will never die. Add grains to input feeds where incoming signals "white out" active grains with overflow, remove grains from zones where signals live below reaction threshold; training is ongoing, with a slow build and slower decay - all grains can also change filter/mask. Remember you have a database backing up this information, use it wisely, this is simply an analysis unit for the database contents.
Assign types to grains: sensor, motor, thinker, memory. Technically memory are sensor neurons with internal feed and feedback. This assignment can be done randomly on a bias - every zone should have at least one of each, MASTER is always a thinker first, and is connected to a simple clock signal; note that this only changes external connections. Zones should each have at least one connection to another zone nearby (sourcing from a root provides this initially), and one zone a bit "further" away in the grouping.
Translate information incoming to this artificial brain into analog or analog-like signals, typically through a vocoder.
This should form a minimal, reactive artifical neuron brain capable of self-learning AI response.
