Multi-memristive synapse with clock-arbitrated weight update

1. A computer-implemented method for providing a multi-memristive synaptic element for a cognitive computing system, said multi-memristive synaptic element comprising an array of memristive devices, said method comprising:

arbitrating, by a programmed hardware processor, a synaptic weight allocation, a related synaptic weight being represented by a synaptic weight variable of said multi-memristive synaptic element,

updating, using said programmed hardware processor, said synaptic weight variable by a delta amount, and

assigning, by the programmed hardware processor, said memristive devices to elements of a clock-like ordered circular list for selecting a particular memristor of said memristive devices requiring to be updated by a deterministic, periodic global clock that points to a different memristor at every clock tick,

running, by said programmed hardware processor, a reset clock, operating independently of said periodic global clock, for triggering a reset operation to be performed on a memristive device;

such that said multi-memristive synaptic element has a larger dynamic range and a more linear conductance response than a single memristor synaptic element; and

wherein a number of said ordered list of elements in said ordered circular list and an increment rate for said ordered circular list relate to each other in such a way that during consecutive iterations of selections at least one sub-element is different.

2. The method according to claim 1 , wherein said memristive devices are organized in a parallel array.

3. The method according to claim 1 , wherein said cognitive computing system is an artificial neural network trained with back-propagation.

4. The method according to claim 3 , wherein an output x j of a post-synaptic neuron is given by

x j =f(Σ(x i ,W ij )), wherein

f is a non-linear function,

x i is an output of a pre-synaptic neuron, and

W ij is said synaptic weight variable.

5. The method according to claim 1 , wherein said amount is positive or negative.

6. The method according to claim 1 , wherein said delta amount ΔW ij of said synaptic weight variable is given by ΔW ij =λ*x i *δ j , wherein

λ is a learning rate,

x i is an output of a pre-synaptic neuron, and

δ j is a back-propagated delta error.

7. The method according to claim 1 , wherein said memristor is a binary memristor or a multilevel memristor.

8. The method according to claim 1 , wherein said reset clock is in the form of an ordered circular list for selecting a particular memristor of said memristive devices for said reset operation at a periodic update frequency.

9. The method according to claim 8 , wherein said reset operations performed at a particular memristor of said memristive devices achieve a gradual change in conductance at said particular memristor.

10. The method according to claim 9 , wherein the update frequencies of the periodic global clock and the reset clock are co-prime to each other.

11. The method according to claim 1 , wherein a total number of memristors of said memristive synaptic element is equal for each memristive synaptic element in said cognitive computing system.

12. The method according to claim 1 , wherein a total number of list elements in said ordered circular list is co-prime to an increment rate for said ordered circular list.

13. A multi-memristive synaptic element for a cognitive computing system, said multi-memristive synaptic element comprising an array of memristive devices, said multi-memristive synaptic element comprising

an arbitration module for arbitrating a synaptic weight allocation, the synaptic weight being represented by a synaptic weight variable of said multi-memristive synaptic element,

an updating unit adapted for updating said synaptic weight variable by a delta amount,

wherein said arbitration module is also adapted for assigning said memristive devices to elements of a clock-like ordered circular list for selecting a particular memristor of said memristive devices requiring to be updated by a deterministic, periodic global clock that points to a different memristor at every clock tick,

a reset clock operating independently of said periodic global clock for triggering a reset operation performed on a memristive device;

such that said multi-memristive synaptic element has a larger dynamic range and a more linear conductance response than a single memristor synaptic element; and

wherein a number of said ordered list of elements in said ordered circular list and an increment rate for said ordered circular list relate to each other in such a way that during consecutive iterations of selections at least one sub-element is different.

14. The multi-memristive synaptic element according to claim 13 , wherein said memristive devices are organized in a parallel array.

15. The multi-memristive synaptic element according to claim 13 , wherein an output x j of a post-synaptic neuron is given by x j =f(Σ(x i ,W ij )), wherein

f is a non-linear function,

x i is an output of a pre-synaptic neuron, and

W ij is said synaptic weight variable.

16. The multi-memristive synaptic element according to claim 14 , wherein said delta amount ΔW ij of said synaptic weight variable is given by ΔW ij =λ*x i *δ j , wherein

λ is a learning rate,

x i is an output of a pre-synaptic neuron, and

δ j is a back-propagated delta error.

17. The multi-memristive synaptic element according to claim 13 , wherein said reset clock is in the form of an ordered circular list for selecting a particular memristor of said memristive devices for said reset operation at a periodic update frequency.

18. The multi-memristive synaptic element according to claim 17 , wherein said reset operations performed at a particular memristor of said memristive devices achieve a gradual change in conductance at said particular memristor.

19. The multi-memristive synaptic element according to claim 18 , wherein the update frequencies of the periodic global clock and the reset clock are co-prime to each other.

20. A computer program product for operating a multi-memristive synaptic element for a cognitive computing system, said multi-memristive synaptic element comprising an array of memristive devices, said computer program product comprising a non-transitory computer readable storage medium having program instructions embodied therewith, said program instructions being executable by one or more computing systems to cause said one or more computing systems to:

arbitrate a synaptic weight allocation, a related synaptic weight being represented by a synaptic weight variable of said multi-memristive synaptic element,

update said synaptic weight variable by a delta amount, and

assign said memristive devices to elements of a clock-like ordered circular list for selecting a particular memristor of said memristive devices required to be updated by a deterministic, periodic global clock that points to a different memristor at every clock tick,

operate a reset clock operating independently of said periodic global clock for triggering a reset operation performed on a memristive device;

such that said multi-memristive synaptic element has a larger dynamic range and a more linear conductance response than a single memristor synaptic element; and

wherein a number of said ordered list of elements in said ordered circular list and an increment rate for said ordered circular list relate to each other in such a way that during consecutive iterations of selections at least one sub-element is different.