Hierarchical temporal memory utilizing nanotechnology
View Patent ↗Methods and systems are presented for constructing biological-scale hierarchically structured cortical statistical memory systems using currently available fabrication technology and meta-stable switching devices. Learning content-addressable memory and statistical random access memory circuits are detailed. Additionally, local and global signal modulation of bottom-up and top-down processing for the initiation and direction of behavior is disclosed.
1. A method for providing a hierarchical temporal memory utilizing nanotechnology, comprising:
mapping at least one arbitrary hierarchical structure for a two-dimensional layout in order to provide for a hierarchical temporal memory;
storing pattern probabilities, wherein said pattern probabilities are fed up and down the hierarchical temporal memory;
utilizing a Flip-Lock cycle to provide Anti-Hebbian learning;
utilizing a Flip-Lock Cycle to provide Hebbian learning, in combination with weight decay;
providing a router hierarchy for generating a feedback signal to control Hebbian and Anti-Hebbian forms of learning;
providing a critical period, wherein during said critical period all neural nodes in an ensemble receive a synaptic update; and
providing an active period, wherein all neural nodes in said router hierarchy compete for a Hebbian update.
2. The method of claim 1 further comprising providing a KBPI subsystem for a prediction override regulation.
3. The method of claim 2 further comprising regulating said prediction override regulation via a Dopa parameter.
4. The method of claim 3 further comprising regulating said Dopa parameter by both global and local rules.
5. The method of claim 3 further comprising regulating said Dopa parameter in proportion to a temporal derivative of said Confidence.
6. The method of claim 1 further comprising measuring a prediction magnitude by post-synaptic current activation.
7. The method of claim 1 further comprising measuring a prediction magnitude by synchronous firing of KBPI neurons.
8. The method of claim 1 further comprising regulating a confidence threshold by adaptively accumulating expectations.
9. The method of claim 1 further comprising increasing a prediction magnitude in an amount proportional to a Dopa parameter.
10. The method of claim 1 further comprising utilizing a decay period for synaptic renormalization.
11. The method of claim 1 further comprising inducing an S state by regulation of a Dopa parameter.
12. The method of claim 11 further comprising utilizing said S state for synaptic repair.
13. The method of claim 12 further comprising:
utilizing a plurality of cycles of decay states followed by S states for synaptic weight conditioning; and
constructing a plurality of meta-KHTM modules from a plurality of smaller KHTM modules.
14. The method of claim 1 wherein at least one arbitrary hierarchical structure is mapped to a two-dimensional layout in order to provide for said hierarchical temporal memory.
15. A method for providing a hierarchical temporal memory utilizing nanotechnology, comprising:
mapping at least one arbitrary hierarchical structure to a two-dimensional layout in order to provide for a hierarchical temporal memory;
storing pattern probabilities, wherein said pattern probabilities are fed up and down the hierarchical temporal memory;
utilizing a Flip-Lock cycle to provide Anti-Hebbian learning;
utilizing a Flip-Lock Cycle to provide Hebbian learning, in combination with weight decay; and
providing a router hierarchy for generating a feedback signal to control Hebbian and Anti-Hebbian forms of the Flip-Lock-cycle;
providing a critical period, wherein during said critical period all neural nodes in an ensemble receive a synaptic update; and
providing an active period, wherein all neural nodes in said router hierarchy compete for a Hebbian update.