Conductance drift corrections in neuromorphic systems based on crossbar array structures

1. A method of operating a neuromorphic system, wherein the method comprises:

applying voltage signals from first signal lines of a control unit across input lines of a crossbar array structure, the crossbar array structure including rows and columns interconnected at first junctions via programmable first electronic devices, the rows including the input lines for applying the voltage signals across the first electronic devices and the columns including output lines for outputting currents; and

correcting each of the output currents obtained at the output lines according to an affine transformation to compensate for temporal conductance variations in the first electronic devices, the correcting including

storing a multiplicative coefficient γ for the affine transformation in programmable second electronic devices in response to receiving second voltage signals from the control unit, the programmable second electronic devices interconnected at second junctions between the output lines and second signal lines of the control unit,

storing an additive parameter β for the affine transformation in programmable third electronic devices in response to receiving third voltage signals from the control unit, the programmable third electronic devices interconnected at third junctions between the output lines and third signal lines of the control unit, and

performing the affine transformation with an integrated circuit based on the stored multiplicative coefficient γ and the stored additive parameter β.

2. The method according to claim 1 , wherein the method further comprises:

computing at least one set of values for the programmable parameters, wherein the second programmable electronic devices and third programmable electronic devices are programmed according to the at least one set of values computed.

3. The method according to claim 2 , wherein:

the columns include at least two sets of output lines outputting at least two sets of output currents, respectively;

at least two sets of values are computed for said programmable parameters; and

programming the second programmable electronic devices and third programmable electronic devices according to the at least two sets of values, to separately correct the at least two sets of output currents according to the at least two sets of values, respectively.

4. The method according to claim 3 , wherein:

the columns include M output lines outputting M output currents, respectively; and

M sets of values are computed for the programmable parameters, for the M output currents to be separately corrected according to the M set of values, respectively.

5. The method according to claim 2 , wherein

the crossbar array structure further includes at least one additional column, which are connected to the rows at fourth junctions via fourth electronic devices, the additional columns including one or more additional output lines for outputting one or more reference currents, and

the at least one set of values are computed according to the one or more reference currents.

6. The method according to claim 5 , wherein the method further comprises programming the fourth junctions for at least two subsets of the fourth electronic devices to exhibit at least two, distinct electrical conductances, respectively, whereby the distinct electrical conductances impact the reference currents outputted by the additional output lines.

7. The method according to claim 2 , wherein

the programming comprises initializing said programmable parameters after having programmed the first electronic devices but prior to computing the at least one set of values for the programmable parameters, whereby the programming is performed according to the at least one set of values computed.

8. The method according to claim 7 , wherein

computing the at least one set of values includes computing several, time-dependent sets of values for said programmable parameters, wherein the programming comprises repeatedly programming the second programmable electronic devices and the third programmable electronic devices according to the time-dependent sets of values.

9. The method according to claim 1 , wherein the method further comprises:

programming the first electronic devices for the first electronic devices to store synaptic weights pertaining to connections to nodes of a single layer of an artificial neural network,

wherein the output currents obtained at the output lines are obtained according to a multiply-accumulate operation based on the voltage signals applied across the input lines and values stored on each of the first electronic devices as programmed on the first electronic devices.

10. The method according to claim 9 , wherein the method further comprises:

computing batch normalization parameters; and

performing a batch normalization of the layer by scaling the multiplicative coefficient γ and the additive parameter β according to the computed batch normalization parameters.

11. A neuromorphic system, comprising:

a crossbar array structure that includes rows and columns interconnected at first junctions via first electronic devices, wherein the rows include input lines for applying voltage signals from first signal lines of a control unit across the first electronic devices and the columns include output lines for outputting currents;

a correction unit connected to the output lines and configured to enable an affine transformation of currents outputted from each of the output lines; and

a control unit configured to

apply the voltage signals across the input lines, and

operate the correction unit to correct each of the output currents obtained at the output lines according to the affine transformation, to compensate for temporal conductance variations in the first electronic devices,

wherein the correction unit is configured to store programmable parameters and includes

programmable second electronic devices interconnected at second junctions between the output lines and second signal lines of the control unit, the second electronic devices configured to store a multiplicative coefficient γ for the affine transformation in response to receiving second voltage signals from the control unit,

programmable third electronic devices interconnected at third junctions between the output lines and third signal lines of the control unit, the third electronic devices configured to store an additive parameter β for the affine transformation in response to receiving a third voltage signals from the control unit, and

an integrated circuit configured to perform the affine transformation based on the stored multiplicative coefficient γ and the stored additive parameter β.

12. The system according to claim 11 , wherein

the control unit is further configured to compute at least one set of values for the programmable parameters and program the correction unit according to the at least one set of values computed, in operation.

13. The system according to claim 12 , wherein

the crossbar array structure further includes one or more additional columns, which are connected to the rows at fourth junctions via fourth electronic devices, the additional columns including one or more additional output lines for outputting one or more reference currents, in operation, and

the control unit is further configured to compute said at least one set of values according to said one or more reference currents.

14. The system according to claim 13 , wherein

the control unit is further configured to program said fourth junctions so as for at least two subsets of the third electronic devices to exhibit at least two, distinct electrical conductances, respectively.

15. The system according to claim 11 , wherein

said columns include at least two sets of output lines for outputting at least two sets of output currents, respectively, and

the control unit is further configured to compute at least two sets of values for said programmable parameters and program the correction unit according to said at least two sets of values, so as to separately correct said at least two sets of output currents according to respective ones of said at least two sets of values, in operation.

16. The system according to claim 11 , wherein

the control unit is further configured to compute several, time-dependent sets of values for said programmable parameters and repeatedly program the correction unit according to said time-dependent sets of values, in operation.