Synaptic weight normalized spiking neuronal networks
Neuronal networks of electronic neurons interconnected via electronic synapses with synaptic weight normalization. The synaptic weights are based on learning rules for the neuronal network, such that a synaptic weight for a synapse determines the effect of a spiking source neuron on a target neuron connected via the synapse. Each synaptic weight is maintained within a predetermined range by performing synaptic weight normalization for neural network stability.
1. A method, comprising:
determining synaptic weights for electronic synapses configured for connecting axons of source electronic neurons to dendrites of target electronic neurons in a spiking neuronal network, wherein the synaptic weights are based on learning rules for the neuronal network, such that a synaptic weight for an electronic synapse determines the effect of a spiking source electronic neuron on a target electronic neuron connected via that electronic synapse; and
dynamically maintaining total synaptic weights in the spiking neuronal network at a predetermined range for neural network stability by adjusting a synaptic weight for at least one electronic synapse;
wherein, for an electronic synapse connected to an electronic neuron, a synaptic weight for said electronic synapse is adjusted based on a total synaptic weight for electronic synapses connected to said electronic neuron.
2. The method of claim 1 , wherein dynamically maintaining total synaptic weights in the spiking neuronal network further comprises:
for each electronic neuron, dynamically maintaining a corresponding total in-coming synaptic weight for said electronic neuron at a predetermined range, wherein said corresponding total in-coming synaptic weight is based on each synaptic weight for each electronic synapse connected to a dendrite of said neuron.
3. The method of claim 2 , wherein:
dynamically maintaining a corresponding total in-coming synaptic weight for an electronic neuron at a predetermined range comprises synaptic weight normalization by adjusting one or more learning rule parameters for a learning rule, wherein a synaptic weight for at least one electronic synapse connected to a dendrite of said electronic neuron is normalized.
4. The method of claim 3 , wherein:
if a corresponding total in-coming synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron; and
if a corresponding total in-coming synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron.
5. The method of claim 1 , wherein dynamically maintaining total synaptic weights in the spiking neuronal network further comprises:
for each electronic neuron, dynamically maintaining a corresponding total out-going synaptic weight for said electronic neuron at a predetermined range, wherein said corresponding total outgoing synaptic weight is based on each synaptic weight for each electronic synapse connected to each axon of said neuron.
6. The method of claim 5 , wherein:
dynamically maintaining a corresponding total out-going synaptic weight for an electronic neuron at a predetermined range comprises synaptic weight normalization by adjusting one or more learning rule parameters for a learning rule, wherein a synaptic weight for at least one electronic synapse connected to an axon of said electronic neuron is normalized.
7. The method of claim 6 , wherein:
if a corresponding total out-going synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron; and
if a corresponding total out-going synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron.
8. The method of claim 1 , wherein dynamically maintaining total synaptic weights in the spiking neuronal network further comprises:
for each electronic neuron, dynamically maintaining a corresponding total in-coming synaptic weight for said electronic neuron at a predetermined range, and dynamically maintaining a corresponding total out-going synaptic weight for said electronic neuron at a predetermined range, wherein said corresponding total in-coming synaptic weight is based on each synaptic weight for each electronic synapse connected to a dendrite of said neuron, and wherein said corresponding total out-going synaptic weight is based on each synaptic weight for each electronic synapse connected to each axon of said neuron.
9. The method of claim 8 , wherein:
dynamically maintaining a corresponding total in-coming synaptic weight for an electronic neuron at a predetermined range, and dynamically maintaining a corresponding total out-going synaptic weight for said electronic neuron at a predetermined range comprises synaptic weight normalization by adjusting one or more learning rule parameters for a learning rule, wherein a synaptic weight for at least one electronic synapse connected to a dendrite of said electronic neuron is normalized.
10. The method of claim 9 , wherein:
if a corresponding total out-going synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron; and
if a corresponding total out-going synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron;
if a corresponding total in-coming synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron; and
if a corresponding total in-coming synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron.
11. The method of claim 4 , further comprising:
utilizing a distributed memory multiprocessor system including multiple processors for simulating a spiking neuronal network, wherein electronic neurons are distributed among local memories of the multiple processors;
wherein a corresponding total in-coming synaptic weight for an electronic neuron is determined by communicating each synaptic weight for each electronic synapse connected to a dendrite of said electronic neuron to a processor corresponding to said electronic neuron, wherein a synaptic weight is communicated said processor based on one or more of the following: on a periodic basis, and with one or more spike messages.
12. The method of claim 4 , wherein:
the neuronal network further comprises a synaptic crossbar array interconnecting the electronic neurons, wherein the crossbar array comprises a plurality of axon paths and a plurality of dendrite paths, wherein the axon paths and the dendrite paths are orthogonal to one another, and wherein each electronic synapse is at a cross-point junction of the crossbar array coupled between a dendrite path and an axon path; and
a total in-coming synaptic weight is determined for each dendrite path, wherein a total in-coming synaptic weight for a dendrite path is based on each synaptic weight for each electronic synapse coupled to said dendrite path.
13. The method of claim 7 , further comprising:
utilizing a distributed memory multiprocessor system including multiple processors for simulating a spiking neuronal network, wherein electronic neurons are distributed among local memories of the multiple processors;
wherein a corresponding total out-going synaptic weight for an electronic neuron is determined by communicating each synaptic weight for each electronic synapse connected to each axon of said electronic neuron to a processor corresponding to said electronic neuron, wherein a synaptic weight is communicated said processor based on one or more of the following: on a periodic basis, and with one or more spike messages.
14. The method of claim 7 , wherein:
the neuronal network further comprises a synaptic crossbar array interconnecting the electronic neurons, wherein the crossbar array comprises a plurality of axon paths and a plurality of dendrite paths, wherein the axon paths and the dendrite paths are orthogonal to one another, and wherein each electronic synapse is at a cross-point junction of the crossbar array coupled between a dendrite path and an axon path; and
a total out-going synaptic weight is determined for each axon path, wherein a total out-going synaptic weight for an axon path is based on each synaptic weight for each electronic synapse coupled to said axon path.
15. The method of claim 10 , further comprising:
utilizing a distributed memory multiprocessor system including multiple processors for simulating a spiking neuronal network, wherein electronic neurons are distributed among local memories of the multiple processors;
wherein a corresponding total out-going synaptic weight for an electronic neuron is determined by communicating each synaptic weight for each electronic synapse connected to each axon of said electronic neuron to a processor corresponding to said electronic neuron, wherein a synaptic weight is communicated said processor based on one or more of the following: on a periodic basis, and with one or more spike messages; and
wherein a corresponding total in-coming synaptic weight for an electronic neuron is determined by communicating each synaptic weight for each electronic synapse connected to a dendrite of said electronic neuron to a processor corresponding to said electronic neuron, wherein a synaptic weight is communicated said processor based on one or more of the following: on a periodic basis, and with one or more spike messages.
16. The method of claim 10 , wherein:
the neuronal network further comprises a synaptic crossbar array interconnecting the electronic neurons, wherein the crossbar array comprises a plurality of axon paths and a plurality of dendrite paths, wherein the axon paths and the dendrite paths are orthogonal to one another, and wherein each electronic synapse is at a cross-point junction of the crossbar array coupled between a dendrite path and an axon path;
a total out-going synaptic weight is determined for each axon path, wherein a total out-going synaptic weight for an axon path is based on each synaptic weight for each electronic synapse coupled to said axon path; and
a total in-coming synaptic weight is determined for each dendrite path, wherein a total in-coming synaptic weight for a dendrite path is based on each synaptic weight for each electronic synapse coupled to said dendrite path.
17. A spiking neuronal network system, comprising:
a synaptic interconnect network for interconnecting electronic neurons, wherein the interconnect network comprises a plurality of axon paths and a plurality of dendrite paths, wherein each electronic synapse is at a cross-point junction of the interconnect network coupled between a dendrite path and an axon path, and wherein each electronic synapse is configured for connecting axons of source electronic neurons to dendrites of target electronic neurons in the crossbar array; and
each electronic synapse having a synaptic weight based on learning rules for the neuronal network, such that a synaptic weight for an electronic synapse determines the effect of a spiking source electronic neuron on a target electronic neuron connected via that electronic synapse;
wherein total synaptic weights in the spiking neuronal network are dynamically maintained at a predetermined range for neural network stability by adjusting a synaptic weight for at least one electronic synapse; and
wherein, for an electronic synapse connected to an electronic neuron, a synaptic weight for said electronic synapse is adjusted based on a total synaptic weight for electronic synapses connected to said electronic neuron.
18. The system of claim 17 , wherein the interconnect network further comprises a synaptic crossbar array interconnecting the electronic neurons, wherein the crossbar array includes said plurality of axon paths and said plurality of dendrite paths, wherein said plurality of axon paths and said plurality of dendrite paths are orthogonal to one another, and wherein each electronic synapse is at a cross-point junction of the crossbar array coupled between a dendrite path and an axon path.
19. A spiking neuronal network system, comprising:
a synaptic interconnect network for interconnecting electronic neurons via electronic synapses, wherein the interconnect network comprises a plurality of axon paths and a plurality of dendrite paths, wherein each electronic synapse is at a cross-point junction of the interconnect network coupled between a dendrite path and an axon path, and wherein each electronic synapse is configured for connecting axons of source electronic neurons to dendrites of target electronic neurons in the crossbar array; and
each electronic synapse having a synaptic weight based on learning rules for the neuronal network, such that a synaptic weight for an electronic synapse determines the effect of a spiking source electronic neuron on a target electronic neuron connected via that electronic synapse;
wherein total synaptic weights in the spiking neuronal network are dynamically maintained at a predetermined range for neural network stability by adjusting a synaptic weight for at least one electronic synapse;
wherein, for an electronic synapse connected to an electronic neuron, a synaptic weight for said electronic synapse is adjusted based on a total synaptic weight for electronic synapses connected to said electronic neuron; and
wherein a corresponding total in-coming synaptic weight and a corresponding out-going synaptic weight for each electronic neuron are maintained at a predetermined range utilizing synaptic weight normalization by adjusting one or more learning rule parameters for a learning rule, wherein a synaptic weight for at least one electronic synapse connected to said electronic neuron is normalized.
20. The system of claim 19 , wherein:
if a corresponding total out-going synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron;
if a corresponding total out-going synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on out going synapse devices electronic synapses connected to an axon of the electronic neuron;
if a corresponding total in-coming synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron; and
if a corresponding total in-coming synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron.
21. A computer program product comprising:
a non-transitory computer usable medium having computer readable program code embodied therewith, wherein the computer readable program code when executed on the computer causes the computer to perform operations comprising:
determining synaptic weights for electronic synapses configured for connecting axons of source electronic neurons to dendrites of target electronic neurons in a spiking neuronal network, wherein the synaptic weights are based on learning rules for the neuronal network, such that a synaptic weight for an electronic synapse determines the effect of a spiking source electronic neuron on a target electronic neuron connected via that electronic synapse; and
dynamically maintaining total synaptic weights in the spiking neuronal network at a predetermined range for neural network stability by adjusting a synaptic weight for at least one electronic synapse;
wherein, for an electronic synapse connected to an electronic neuron, a synaptic weight for said electronic synapse is adjusted based on a total synaptic weight for electronic synapses connected to said electronic neuron.
22. The computer program product of claim 21 , wherein the program code further causes the computer to perform operations further comprising:
dynamically maintaining total synaptic weights in the spiking neuronal network by, for each electronic neuron, dynamically maintaining a corresponding total in-coming synaptic weight for said electronic neuron at a predetermined range, wherein said corresponding total in-coming synaptic weight is based on each synaptic weight for each electronic synapse connected to a dendrite of said neuron; and
dynamically maintaining a corresponding total in-coming synaptic weight for an electronic neuron at a predetermined range comprises synaptic weight normalization by adjusting one or more learning rule parameters for a learning rule, wherein a synaptic weight for at least one electronic synapse connected to a dendrite of said electronic neuron is normalized;
wherein, if a corresponding total in-coming synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron; and
wherein, if a corresponding total in-coming synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron.
23. The computer program product of claim 21 , wherein the program code further causes the computer to perform operations further comprising:
dynamically maintaining total synaptic weights in the spiking neuronal network by, for each electronic neuron, dynamically maintaining a corresponding total out-going synaptic weight for said electronic neuron at a predetermined range, wherein said corresponding total out-going synaptic weight is based on each synaptic weight for each electronic synapse connected to each axon of said neuron; and
dynamically maintaining a corresponding total out-going synaptic weight for an electronic neuron at a predetermined range comprises synaptic weight normalization by adjusting one or more learning rule parameters for a learning rule, wherein a synaptic weight for at least one electronic synapse connected to an axon of said electronic neuron is normalized;
wherein, if a corresponding total out-going synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapse connected to an axon of the electronic neuron; and
wherein, if a corresponding total out-going synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron.
24. The computer program product of claim 21 , wherein the program code further causes the computer to perform operations further comprising:
dynamically maintaining total synaptic weights in the spiking neuronal network by, for each electronic neuron, dynamically maintaining a corresponding total in-coming synaptic weight and a corresponding total out-going synaptic weight for said electronic neuron at a predetermined range, wherein said corresponding total in-coming synaptic weight is based on each synaptic weight for each electronic synapse connected to a dendrite of said neuron, and wherein said corresponding total out-going synaptic weight is based on each synaptic weight for each electronic synapse connected to each axon of said neuron;
dynamically maintaining a corresponding total in-coming synaptic weight and a corresponding total out-going synaptic weight for an electronic neuron at a predetermined range utilizing synaptic weight normalization by adjusting one or more learning rule parameters for a learning rule, wherein a synaptic weight for at least one electronic synapse connected to an axon of said electronic neuron is normalized;
wherein, if a corresponding total out-going synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron;
wherein, if a corresponding total out-going synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to an axon of the electronic neuron;
wherein, if a corresponding total in-coming synaptic weight for an electronic neuron is greater than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule decreases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron; and
wherein, if a corresponding total in-coming synaptic weight for an electronic neuron is less than a predetermined range, one or more learning rule parameters for a learning rule are adjusted such that the learning rule increases synaptic potentiation on electronic synapses connected to a dendrite of the electronic neuron.
25. The computer program product of claim 21 , wherein:
the computer comprises a distributed memory multiprocessor system including multiple processors for simulating a spiking neuronal network, wherein electronic neurons are distributed among local memories of the multiple processors;
the program code further causes the computer to perform operations wherein a corresponding total in-coming synaptic weight for an electronic neuron is determined by communicating each synaptic weight for each electronic synapse connected to a dendrite of said electronic neuron to a processor corresponding to said electronic neuron, wherein a synaptic weight is communicated said processor based on one or more of the following: on a periodic basis, and with one or more spike messages.