DETAILED ACTION
Response to Amendment
The amendment filed on 24 December 2025 has been entered.
Claims 1, 3-12 are pending.
Claims 1, 6, 8, 10 are amended.
Claims 13-18 are new.
Claims 1, 3-18 will be pending.
Applicant’s amendments to the Claims have overcome each and every objection and rejection under 35 USC 112(b) and 35 USC 101, previously set forth in the Non-Final Office Action mailed 24 September 2025. Claim interpretations under 35 U.S.C. 112(f), previously set forth in the Non-Final Office Action mailed 24 September 2025, have been withdrawn in view of Applicant’s amendments to the Claims.
Response to Arguments
Applicant's arguments filed on 24 December 2025 have been fully considered, but they are not persuasive.
Applicant’s remarks, regarding the rejections of claims under 35 USC 103, have been fully considered.
Applicant respectfully submits that the amended independent Claims 1, 6, 8, and 10 now recite "wherein the target threshold is different from a real threshold of each neuron," which distinguishes over the cited prior art combination and renders the claims allowable. Applicant submits Claims 3-5, 7, 9, and 11-12 are allowable for the same reasons as their respective independent claims from which they depend.
Applicant’s arguments have been considered, but are moot because the new ground of rejection does not rely on any reference applied in the prior rejection of record for any teaching or matter specifically challenged in the argument.
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Objections
Claim 5 is objected to because of the following informalities: “the each neuron” in line 2 should be “the each neuron of the plurality of neurons”. Appropriate correction is required.
Claim 7 is objected to because of the following informalities: “each neuron” in line 1 should be “each neuron of the plurality of neurons”. Appropriate correction is required.
Claim 9 is objected to because of the following informalities: “the input signals” in line 2 should be “the accumulated input signals”. Appropriate correction is required.
Claim 9 is objected to because of the following informalities: “the neuron” in line 3 should be “the neuron of the plurality of neurons”. Appropriate correction is required.
Claim 10 is objected to because of the following informalities: “the spike” in line 4 should be “the output spike”. Appropriate correction is required.
Claim 10 is objected to because of the following informalities: “each neuron” in line 6 should be “each neuron of the plurality of neurons”. Appropriate correction is required.
Claim 10 is objected to because of the following informalities: “the amount of the accumulated input signal” in line 7 should be “the amount of the accumulated input signals”. Appropriate correction is required.
Claim 10 is objected to because of the following informalities: “the accumulation unit” in line 8 should be “an accumulation unit”. Appropriate correction is required.
Claim 10 is objected to because of the following informalities: “the amount of a target threshold” in line 8 should be “the same amount of a target threshold”. Appropriate correction is required.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
Claims 1, 3-18 are rejected under 35 U.S.C. 103 as being unpatentable over Querlioz et al. (“Immunity to Device Variations in a Spiking Neural Network With Memristive Nanodevices”; hereinafter “Querlioz”), in view of Kesari et al. (“Enabling Homeostasis using Temporal Decay Mechanisms in Spiking CNNs Trained with Unsupervised Spike Timing Dependent Plasticity”, conference end 24 July 2020; hereinafter “Kesari”), and further in view of Oh et al. (U.S. Pre-Grant Publication No. 20200160146, hereinafter ‘Oh’).
As per Claim 1, Querlioz teaches a method for compensating a difference and/or a variation of a neuron threshold for a firing in a spiking neural network comprising a plurality of neurons, the method comprising (Querlioz, Page 293 Section III B: “We now study the impact of the variability of the CMOS neurons. The impact of the variability of their threshold is seen in Fig. 10. We can see that without homeostasis the impact of the variability is dramatic. The more excitable neurons (the ones with lower thresholds) spike predominantly and the other neurons do not specialize efficiently. In a typical simulation with a threshold variation of 25%, the most excitable neuron spikes 51% of the time, and in a simulation with a threshold variation of 50%, it spikes 91% of the time. Homeostasis, however, fully compensates this issue (all the neurons spike between 1.5 and 3% of the time, most neurons spiking around 2%). And the same recognition rate as without threshold variation is achieved as evidenced in Fig. 10. Homeostasis appears as particularly valuable in this application. This also stresses the importance to evaluate the robustness of the system not only to nanodevices’ variations, but also to the CMOS’.”
Examiner notes that here, Querlioz teaches that their method of “homeostasis” is effective to “fully compensate” the issue of “variability of the CMOS neurons”).
adjusting an effective threshold for a next firing after a firing of each neuron of the plurality of neurons (Querlioz, Page 291 Section II B 3: “3) Homeostasis: A final issue for the architecture is the adjustment of the neurons’ threshold … Regularly, the threshold of the neuron is increased if the average activity of the neuron is above the target, and decreased if it is below.”
Here, Querlioz teaches that the threshold is adjusted after firing (“activity of the neuron”) for a next firing (“activity of the neuron”)).
wherein adjusting an effective threshold of each neuron of the plurality of neurons comprises lowering a membrane potential of each neuron of the plurality of neurons by an amount of the [target] threshold after the firing of each neuron of the plurality of neurons (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the “membrane potential” is lowered by the full amount of the threshold (“reset to zero”) after the firing of a neuron (“declares a spike”)).
wherein an amount of charging potential for a next spike generation equals to the [target] threshold (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the “threshold” is the “membrane potential” at which a spike is generated.)
However, Querlioz does not teach to be the same as a target threshold in each neuron of the plurality of neurons; lowering a membrane potential of each neuron by an amount of the target threshold after the firing of each neuron of the plurality of neurons, and wherein the target threshold is different from a real threshold of each neuron of the plurality of neurons, and wherein the spiking neural network is implemented in hardware comprising membrane capacitors and discharge transistors.
Kesari teaches to be the same as a target threshold in each neuron of the plurality of neurons; lowering a membrane potential of each neuron of the plurality of neurons by an amount of the target threshold after the firing of each neuron of the plurality of neurons (Kesari, Page 4 Section III B: “A widely adopted homeostasis enabling method in fully connected SNNs is adaptive threshold, wherein the neuron’s threshold is increased based on its activity. This enables other neurons learn a receptive field that is different. In spiking CNNs, the adaptive threshold is defined per activation map, with all neurons in an activation map bearing the same threshold [14]. As the threshold of the activation map is dictated by its activity, the threshold value at any time instant is indicative of its past activity.”
Here, Kesari teaches a same target threshold in each neuron, in that the “target threshold” is one that equals the thresholds of all the others in the activation map. Furthermore, if all neurons have the same target threshold, then when they are reset to zero (as taught above by Querlioz), then their potentials are all lowered by the amount of the target threshold.)
Kesari is analogous art because it is in the field of endeavor of homeostasis in spiking neural networks. It would have been obvious before the effective filing date of the claimed invention to combine the spiking neural network of Querlioz with the common target threshold for an activation map of Kesari. One of ordinary skill in the art would have been motivated to do so in order to achieve homeostasis and balance neuronal activity, which is similar to the goal of Querlioz above (Kesari, End of Page 6 to Page 7: “First, we quantitatively show that the offset and weight decay mechanisms, coupled with adaptive threshold, lead to homeostatic behavior. In order to do so, we fit a Gaussian distribution to the thresholds across the maps. We observe that by modulating the weights and offset decay constants, we are able to obtain a desired amount of homeostasis.”)
Oh teaches wherein the target threshold is different from a real threshold of each neuron of the plurality of neurons ([0026] is different from a real threshold of each neuron of the plurality of neurons Each of the neurons 131 may compare a sum signal in which the operation signals of the synapses 121 are accumulated with a wherein the target threshold threshold signal (that is, a reference signal) and generate an output spike signal when the sum signal is greater than the threshold signal (that is, fire of a neuron).), and
wherein the spiking neural network is implemented in hardware comprising membrane capacitors and discharge transistors ([0037] The wherein the spiking neural network is implemented in hardware spike neural network circuit 100 may further include other capacitors in which charges are accumulated by currents output from other synapses. The capacitor Cm may be referred to as a comprising membrane capacitors membrane capacitor or a membrane.; [0038] The spike neural network circuit 100_1 may include and discharge transistors a transistor MN1 which discharges the charges accumulated in the capacitor Cm depending on a leakage signal. The transistor MN1 may receive the leakage signal through a gate terminal. The transistor MN1 may be connected between the capacitor Cm and the power supply voltage GND. The transistor MN1 may be connected to the capacitor Cm in parallel. The transistor MN1 may control the rate (speed) at which operation signals output from the first to third synapses 121_1 to 121_3 are accumulated in the capacitor Cm. A voltage of the leakage signal may be pre-defined. The transistor MN1 is illustrated as being an NMOS in FIG. 2, but may be implemented using a PMOS, an NMOS, or a combination of the PMOS and the NMOS.).
Querlioz, Kesari, and Oh are considered to be analogous to the claimed invention because they are in the same field of endeavor of homeostasis in spiking neural networks. In view of the teachings of Querlioz and Kesari, it would have been obvious for a person of ordinary skill in the art to apply the teachings of Oh to Querlioz before the effective filing date of the claimed invention in order to reduce power consumption by the spike neural network circuit (cf. Oh, [0078] A spike neural network circuit according to an embodiment of the inventive concept may include a comparator operated by a conditional bias current. Accordingly, the power consumption by the spike neural network circuit may be reduced.).
As per Claim 3, the combination of Querlioz, Kesari, and Oh, teaches the method of claim 1 as well as target threshold (see Kesari in rejection to claim 1).
Querlioz teaches a neural network is implemented in a hardware (Querlioz, Page 2 Section II: “We first introduce the architecture that we propose for our classifier system. The CMOS input and output “neurons” are connected by the nanodevices that act as synapses.”)
wherein the target threshold is lower than a designed threshold of a neuron (Querlioz, Page 4 Right Column: “Regularly, the threshold of the neuron is increased if the average activity of the neuron is above the target, and decreased if it is below.” Here, the target threshold is lower, so the effective threshold must be decreased.)
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 4, the combination of Querlioz, Kesari, and Oh, teaches the method of claim 1 as well as target threshold (see Kesari in rejection to claim 1).
Querlioz teaches wherein an effective threshold for a second and all subsequent firings is set equal to the target threshold. (Querlioz, Page 291 Section II B 3: “3) Homeostasis: A final issue for the architecture is the adjustment of the neurons’ threshold … Regularly, the threshold of the neuron is increased if the average activity of the neuron is above the target, and decreased if it is below.” Here, Querlioz teaches that the threshold is adjusted until activity reaches a homeostatic target, and Examiner notes that in this case when the activity reaches the homeostatic target, the threshold will be effective for subsequent firings.)
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 5, the combination of Querlioz, Kesari, and Oh, teaches the method of claim 1 as well as target threshold (see Kesari in rejection to claim 1).
Querlioz teaches wherein the target threshold is adjusted to be less than a real threshold of the each neuron. (Querlioz, Page 4 Right Column: “Regularly, the threshold of the neuron is increased if the average activity of the neuron is above the target, and decreased if it is below.” Here, the target threshold is lower, so the effective threshold must be decreased.)
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 6, Querlioz teaches a method for discharging a membrane potential after a firing of a neuron in a spiking neural network comprising a plurality of neurons, the method comprising (Querlioz, Page 291 Section II B 1: “The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”)
discharging a [same] amount of a membrane potential for each neuron of the plurality of neurons after the firing (Querlioz, Page 291 Section II B 1: “The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”)
wherein discharging the same amount of a membrane potential for each neuron of the plurality of neurons after the firing is discharging an amount of a [target] threshold for each neuron of the plurality of neurons after the firing (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the “membrane potential” is lowered by the full amount of the threshold (“reset to zero”) after the firing of a neuron (“declares a spike”)).
wherein an amount of charging potential for a next spike generation equals to the [target] threshold (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the “threshold” is the “membrane potential” at which a spike is generated.)
However, Querlioz does not teach discharging a same amount of a membrane potential; discharging an amount of a target threshold, wherein the target threshold is different from a real threshold of each neuron of the plurality of neurons, and wherein the spiking neural network comprises hardware membrane capacitors that are discharged by transistor circuits.
Kesari teaches discharging a same amount of a membrane potential; discharging an amount of a target threshold (Kesari, Page 4 Section III B: “A widely adopted homeostasis enabling method in fully connected SNNs is adaptive threshold, wherein the neuron’s threshold is increased based on its activity. This enables other neurons learn a receptive field that is different. In spiking CNNs, the adaptive threshold is defined per activation map, with all neurons in an activation map bearing the same threshold [14]. As the threshold of the activation map is dictated by its activity, the threshold value at any time instant is indicative of its past activity.”
Here, Kesari teaches a same target threshold in each neuron, in that the “target threshold” is one that equals the thresholds of all the others in the activation map. Furthermore, if all neurons have the same target threshold, then when they are reset to zero (as taught above by Querlioz), then their potentials are all lowered by the amount of the target threshold, and thus they all discharge the same amount of membrane potential.)
Kesari is analogous art because it is in the field of endeavor of homeostasis in spiking neural networks. It would have been obvious before the effective filing date of the claimed invention to combine the spiking neural network of Querlioz with the common target threshold for an activation map of Kesari. One of ordinary skill in the art would have been motivated to do so in order to achieve homeostasis and balance neuronal activity, which is similar to the goal of Querlioz above (Kesari, End of Page 6 to Page 7: “First, we quantitatively show that the offset and weight decay mechanisms, coupled with adaptive threshold, lead to homeostatic behavior. In order to do so, we fit a Gaussian distribution to the thresholds across the maps. We observe that by modulating the weights and offset decay constants, we are able to obtain a desired amount of homeostasis.”)
Oh teaches wherein the target threshold is different from a real threshold of each neuron of the plurality of neurons ([0026] is different from a real threshold of each neuron of the plurality of neurons Each of the neurons 131 may compare a sum signal in which the operation signals of the synapses 121 are accumulated with a wherein the target threshold threshold signal (that is, a reference signal) and generate an output spike signal when the sum signal is greater than the threshold signal (that is, fire of a neuron).), and
wherein the spiking neural network comprises hardware membrane capacitors that are discharged by transistor circuits ([0037] The wherein the spiking neural network comprises spike neural network circuit 100 may further include other capacitors in which charges are accumulated by currents output from other synapses. The capacitor Cm may be referred to as a hardware membrane capacitors membrane capacitor or a membrane.; [0038] The spike neural network circuit 100_1 may include that are discharged by transistor circuits a transistor MN1 which discharges the charges accumulated in the capacitor Cm depending on a leakage signal. The transistor MN1 may receive the leakage signal through a gate terminal. The transistor MN1 may be connected between the capacitor Cm and the power supply voltage GND. The transistor MN1 may be connected to the capacitor Cm in parallel. The transistor MN1 may control the rate (speed) at which operation signals output from the first to third synapses 121_1 to 121_3 are accumulated in the capacitor Cm. A voltage of the leakage signal may be pre-defined. The transistor MN1 is illustrated as being an NMOS in FIG. 2, but may be implemented using a PMOS, an NMOS, or a combination of the PMOS and the NMOS.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 7, the combination of Querlioz, Kesari, and Oh, teaches the method of claim 6. Querlioz teaches wherein each neuron fires to generate a spike when a membrane potential of a time integration and accumulation of an input signal exceed a threshold. (Querlioz, Page 291 Section II B: “1) Output Neurons’ Dynamics: Exploiting the devices requires connecting them to processing units — silicon neurons able to process and generate spikes in a bioinspired manner by integration of their input. We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type, which is meant to solve the simple following equation:
PNG
media_image1.png
46
152
media_image1.png
Greyscale
where τ is a leak time constant, and g and γ are constants … The neuron declares a spike if X reaches a given threshold Xth , in which case X is reset to zero.”)
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 8, Querlioz teaches a resetting method for a neuron for a next firing after a firing of a neuron in a neural network comprising a plurality of neurons, the resetting comprising (Querlioz, Page 291 Section II B 1: “The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”)
reducing an amount of accumulated input signals by a [same] amount for each neuron of the plurality of neurons (Querlioz, Page 291 Section II B 1: “The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”)
wherein the reducing the amount of accumulated input signals by the [same] amount for each neuron of the plurality of neurons is reducing an amount of a [target] threshold for each neuron of the plurality of neurons (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the accumulated input signals (“integrate”, “membrane potential”) is lowered by the full amount of the threshold (“reset to zero”) after the firing of a neuron (“declares a spike”)).
wherein an amount of charging potential for a next spike generation equals to the [target] threshold (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the “threshold” is the “membrane potential” at which a spike is generated.)
However, Querlioz does not teach reducing an amount of accumulated input signals by a same amount for each neuron of the plurality of neurons; reducing an amount of a target threshold, wherein the target threshold is different from a real threshold of each neuron of the plurality of neurons, and wherein the neural network is implemented in hardware with accumulation circuits.
Kesari teaches reducing an amount of accumulated input signals by a same amount for each neuron of the plurality of neurons; reducing an amount of a target threshold (Kesari, Page 4 Section III B: “A widely adopted homeostasis enabling method in fully connected SNNs is adaptive threshold, wherein the neuron’s threshold is increased based on its activity. This enables other neurons learn a receptive field that is different. In spiking CNNs, the adaptive threshold is defined per activation map, with all neurons in an activation map bearing the same threshold [14]. As the threshold of the activation map is dictated by its activity, the threshold value at any time instant is indicative of its past activity.”
Here, Kesari teaches a same target threshold in each neuron, in that the “target threshold” is one that equals the thresholds of all the others in the activation map. Furthermore, if all neurons have the same target threshold, then when they are reset to zero (as taught above by Querlioz), then their potentials are all lowered by the amount of the target threshold, and thus they all discharge the same amount of accumulated input signals.)
Kesari is analogous art because it is in the field of endeavor of homeostasis in spiking neural networks. It would have been obvious before the effective filing date of the claimed invention to combine the spiking neural network of Querlioz with the common target threshold for an activation map of Kesari. One of ordinary skill in the art would have been motivated to do so in order to achieve homeostasis and balance neuronal activity, which is similar to the goal of Querlioz above (Kesari, End of Page 6 to Page 7: “First, we quantitatively show that the offset and weight decay mechanisms, coupled with adaptive threshold, lead to homeostatic behavior. In order to do so, we fit a Gaussian distribution to the thresholds across the maps. We observe that by modulating the weights and offset decay constants, we are able to obtain a desired amount of homeostasis.”)
Oh teaches wherein the target threshold is different from a real threshold of each neuron of the plurality of neurons ([0026] is different from a real threshold of each neuron of the plurality of neurons Each of the neurons 131 may compare a sum signal in which the operation signals of the synapses 121 are accumulated with a wherein the target threshold threshold signal (that is, a reference signal) and generate an output spike signal when the sum signal is greater than the threshold signal (that is, fire of a neuron).), and
wherein the neural network is implemented in hardware with accumulation circuits ([0037] The wherein the neural network is implemented in hardware spike neural network circuit 100 may further include other capacitors in which charges are accumulated by currents output from other synapses. The capacitor Cm may be referred to as a membrane capacitor or a membrane.; [0038] The spike neural network circuit 100_1 may with accumulation circuits include a transistor MN1 which discharges the charges accumulated in the capacitor Cm depending on a leakage signal. The transistor MN1 may receive the leakage signal through a gate terminal. The transistor MN1 may be connected between the capacitor Cm and the power supply voltage GND. The transistor MN1 may be connected to the capacitor Cm in parallel. The transistor MN1 may control the rate (speed) at which operation signals output from the first to third synapses 121_1 to 121_3 are accumulated in the capacitor Cm. A voltage of the leakage signal may be pre-defined. The transistor MN1 is illustrated as being an NMOS in FIG. 2, but may be implemented using a PMOS, an NMOS, or a combination of the PMOS and the NMOS.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 9, the combination of Querlioz, Kesari, and Oh, teaches the resetting method of claim 8.
Querlioz teaches wherein the accumulated input signals are obtained by time integration of the input signals, and wherein the neuron fires to generate a spike when the accumulated input signals exceed a threshold. (Querlioz, Page 291 Section II B: “1) Output Neurons’ Dynamics: Exploiting the devices requires connecting them to processing units — silicon neurons able to process and generate spikes in a bioinspired manner by integration of their input. We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type, which is meant to solve the simple following equation:
PNG
media_image1.png
46
152
media_image1.png
Greyscale
where τ is a leak time constant, and g and γ are constants … The neuron declares a spike if X reaches a given threshold Xth , in which case X is reset to zero.”)
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 10, Querlioz teaches a spiking neural network comprising a plurality of neurons (Querlioz, Abstract: “We propose a novel neural network-based computing paradigm, which exploits their specific physics, and which has virtual immunity to their variability. Memristive devices are used as synapses in a spiking neural network performing unsupervised learning.”)
wherein the plurality of neurons time-integrate and accumulate an input signal and fire an output spike when the accumulated input signal exceeds a threshold (Querlioz, Page 291 Section II B: “1) Output Neurons’ Dynamics: Exploiting the devices requires connecting them to processing units — silicon neurons able to process and generate spikes in a bioinspired manner by integration of their input. We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type, which is meant to solve the simple following equation:
PNG
media_image1.png
46
152
media_image1.png
Greyscale
where τ is a leak time constant, and g and γ are constants … The neuron declares a spike if X reaches a given threshold Xth , in which case X is reset to zero.”)
wherein after firing the spike, an amount of accumulated input signals is reduced by a [same] amount of a [target] threshold for firing a next spike (Querlioz, Page 291 Section II B 1: “The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”)
wherein each neuron comprises: a threshold adjusting unit reducing the amount of the accumulated input signal of the accumulation unit by the amount of a [target] threshold in response to the output spike of a firing unit (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the accumulated input signals (“integrate”, “membrane potential”) is lowered by the full amount of the threshold (“reset to zero”) after the firing of a neuron (“declares a spike”)).
Querlioz also teaches a threshold adjusting unit (Page 288: “using Complementary Metal–Oxide Semiconductor (CMOS) circuits to model its spiking neurons and synapses.”)
wherein an amount of charging potential for a next spike generation equals to the [target] threshold (Querlioz, Page 291 Section II B 1: “We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type … The neuron declares a spike if X reaches a given threshold Xth, in which case X is reset to zero.”
Here, Querlioz teaches that the “threshold” is the “membrane potential” at which a spike is generated.)
However, Querlioz does not teach reduced by a same amount of a target threshold, wherein the target threshold is different from a real threshold of each neuron, and wherein the threshold adjusting unit comprises a discharge transistor connected to a membrane capacitor.
Kesari teaches reduced by a same amount of a target threshold (Kesari, Page 4 Section III B: “A widely adopted homeostasis enabling method in fully connected SNNs is adaptive threshold, wherein the neuron’s threshold is increased based on its activity. This enables other neurons learn a receptive field that is different. In spiking CNNs, the adaptive threshold is defined per activation map, with all neurons in an activation map bearing the same threshold [14]. As the threshold of the activation map is dictated by its activity, the threshold value at any time instant is indicative of its past activity.”
Here, Kesari teaches a same target threshold in each neuron, in that the “target threshold” is one that equals the thresholds of all the others in the activation map. Furthermore, if all neurons have the same target threshold, then when they are reset to zero (as taught above by Querlioz), then their potentials are all lowered by the amount of the target threshold, and thus they all discharge the same amount of accumulated input signals.)
Kesari is analogous art because it is in the field of endeavor of homeostasis in spiking neural networks. It would have been obvious before the effective filing date of the claimed invention to combine the spiking neural network of Querlioz with the common target threshold for an activation map of Kesari. One of ordinary skill in the art would have been motivated to do so in order to achieve homeostasis and balance neuronal activity, which is similar to the goal of Querlioz above (Kesari, End of Page 6 to Page 7: “First, we quantitatively show that the offset and weight decay mechanisms, coupled with adaptive threshold, lead to homeostatic behavior. In order to do so, we fit a Gaussian distribution to the thresholds across the maps. We observe that by modulating the weights and offset decay constants, we are able to obtain a desired amount of homeostasis.”)
Oh teaches wherein the target threshold is different from a real threshold of each neuron ([0026] is different from a real threshold of each neuron Each of the neurons 131 may compare a sum signal in which the operation signals of the synapses 121 are accumulated with a wherein the target threshold threshold signal (that is, a reference signal) and generate an output spike signal when the sum signal is greater than the threshold signal (that is, fire of a neuron).), and
wherein the threshold adjusting unit comprises a discharge transistor connected to a membrane capacitor ([0037] The wherein the threshold adjusting unit spike neural network circuit 100 may further include other capacitors in which charges are accumulated by currents output from other synapses. The capacitor Cm may be referred to as a membrane capacitor or a membrane.; [0038] The spike neural network circuit 100_1 may include comprises a discharge transistor a transistor MN1 which discharges the charges connected to a membrane capacitor accumulated in the capacitor Cm depending on a leakage signal. The transistor MN1 may receive the leakage signal through a gate terminal. The transistor MN1 may be connected between the capacitor Cm and the power supply voltage GND. The transistor MN1 may be connected to the capacitor Cm in parallel. The transistor MN1 may control the rate (speed) at which operation signals output from the first to third synapses 121_1 to 121_3 are accumulated in the capacitor Cm. A voltage of the leakage signal may be pre-defined. The transistor MN1 is illustrated as being an NMOS in FIG. 2, but may be implemented using a PMOS, an NMOS, or a combination of the PMOS and the NMOS.).
combinable
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 11, the combination of Querlioz, Kesari, and Oh, teaches the spiking neural network of claim 10.
Querlioz teaches wherein each neuron comprises: a processor coupled to a memory storing instructions to permit the processor to function as: an accumulation unit accumulating an input signal by time-integrating; and a firing unit firing an output signal when the accumulated input signal of the accumulation unit exceeds a threshold. (Querlioz, Page 291 Section II B: “1) Output Neurons’ Dynamics: Exploiting the devices requires connecting them to processing units — silicon neurons able to process and generate spikes in a bioinspired manner by integration of their input. We call X the state variable (a current or a voltage, equivalent of the biological “membrane potential”) of the neuron (expressed in normalized unit where the maximum value of the state variable X is 1). Neurons are leaky integrate-and-fire type, which is meant to solve the simple following equation:
PNG
media_image1.png
46
152
media_image1.png
Greyscale
where τ is a leak time constant, and g and γ are constants … The neuron declares a spike if X reaches a given threshold Xth , in which case X is reset to zero.”
Querlioz also teaches a processor comprising units in Page 288 Top Right: “using Complementary Metal–Oxide Semiconductor (CMOS) circuits to model its spiking neurons and synapses” and Page 289 Section II: “The CMOS input and output “neurons” are connected by the nanodevices that act as synapses. It is natural to lay out the nanodevices in the widely studied crossbar as illustrated on Fig. 1, where CMOS silicon neurons and their associated synaptic driving circuitry are the dots, the squares being the nanodevices. The synapses indeed act as adaptive resistors.”)
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
As per Claim 12, the combination of Querlioz, Kesari, and Oh, teaches the spiking neural network of claim 11.
Querlioz teaches wherein each neuron, after firing a first output spike, fires subsequent output spikes when accumulated input signals in the accumulation unit exceed the target threshold. (Querlioz, Page 291 Section II B 3: “3) Homeostasis: A final issue for the architecture is the adjustment of the neurons’ threshold … Regularly, the threshold of the neuron is increased if the average activity of the neuron is above the target, and decreased if it is below.” Here, Querlioz teaches that the threshold is adjusted until activity reaches a homeostatic target, and Examiner notes that in this case when the activity reaches the homeostatic target, the threshold will be effective for subsequent firings.)
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 13, Querlioz, as modified by Kesari and Oh, teaches The method of claim 1.
Oh teaches wherein the membrane capacitors have capacitance values that vary based on manufacturing process variations, and wherein discharging by the target threshold compensates for the capacitance variations ([0037] The spike neural network circuit 100 may further include other capacitors in which charges are accumulated by currents output from other synapses. The capacitor Cm may be referred to as a membrane capacitor or a membrane.; [0054] The capacitor Cq and the transistors MN9 to MN11 and MP9 may configure a quiescence adjustment circuit 134_1 b which lowers the voltage Vm of the membrane signal to the power supply voltage GND. The quiescence adjustment circuit 134_1 b may wherein discharging by the target threshold compensates for the capacitance variations adjust an interval in which the membrane signal is deactivated or an interval in which the output spike signal is deactivated. A quiescence of the neuron 131_1 may represent a duration (time) during which the voltage Vm of the membrane signal is driven or maintained with the power supply voltage GND corresponding to a reset, or a duration during which the output spike signal is activated and then deactivated. The quiescence may be adjusted based on the second bias signal, the transistor MN10, and wherein the membrane capacitors have capacitance values that vary based on manufacturing process variations a capacity of the capacitor Cq. Even when an input spike signal is activated and operation results are output from the synapses 121 in the quiescence, since the voltage Vm of the membrane signal is maintained as the power supply voltage GND, the operation results may be ignored.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 14, Querlioz, as modified by Kesari and Oh, teaches The method of claim 1.
Oh teaches wherein the discharge transistors are NMOS transistors having gates connected to neuron output terminals and drains connected to the membrane capacitors ([0038] The spike neural network circuit 100_1 may include a transistor MN1 and drains connected to the membrane capacitors which discharges the charges accumulated in the capacitor Cm depending on a leakage signal. The transistor MN1 may having gates connected to neuron output terminals receive the leakage signal through a gate terminal. The transistor MN1 may be connected between the capacitor Cm and the power supply voltage GND. The transistor MN1 may be connected to the capacitor Cm in parallel. The transistor MN1 may control the rate (speed) at which operation signals output from the first to third synapses 121_1 to 121_3 are accumulated in the capacitor Cm. A voltage of the leakage signal may be pre-defined. The wherein the discharge transistors are NMOS transistors transistor MN1 is illustrated as being an NMOS in FIG. 2, but may be implemented using a PMOS, an NMOS, or a combination of the PMOS and the NMOS.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 15, Querlioz, as modified by Kesari and Oh, teaches The method of claim 1.
Oh teaches wherein an excess membrane potential above the real threshold is preserved after discharging by the target threshold amount ([0026] Each of the neurons 131 may compare a sum signal in which the wherein an excess membrane potential above the real threshold is preserved after discharging by the target threshold amount operation signals of the synapses 121 are accumulated with a threshold signal (that is, a reference signal) and generate an output spike signal when the sum signal is greater than the threshold signal (that is, fire of a neuron).; [0037] The spike neural network circuit 100_1 may include a capacitor Cm in which charges are accumulated (integrated) by the first to third operation signals (currents) output from the first to third synapses 121_1 to 121_3. A first end of the capacitor Cm may be connected to the first to third synapses 121_1 to 121_3, and a second end of the capacitor Cm may be connected to a power supply voltage (ground voltage) GND. The capacitor Cm may be charged by currents output from the first to third synapses 121_1 to 121_3 and corresponding to the first to third weights. The voltage Vm of the capacitor Cm is the voltage Vm of the membrane signal, and may be a value in which currents output from the first to third synapses 121_1 to 121_3 are accumulated. The voltage Vm of the capacitor Cm may be a value determined by the first to third weights output from the first to third synapses 121_1 to 121_3 to the first to third input spike signals. The voltage Vm of the capacitor Cm may be provided to the neuron 131_1. The number of synapses connected to the capacitor Cm through the transmission line is illustrated as being 3 in FIG. 2.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 16, Querlioz, as modified by Kesari and Oh, teaches The method of claim 1.
Oh teaches wherein the hardware implementation eliminates need for negative voltage supplies by setting the target threshold lower than real thresholds of all neurons ([0034] The neuron 131_1 may include a comparator 132_1 which compares a membrane signal (a sum signal) in which operation signals output from the first to three synapses 121_1 to 121_3 are combined and a threshold signal. The membrane signal may be generated based on the operation signals. The comparator 132_1 may compare a voltage Vm of the membrane signal with a voltage Vth of the threshold signal. The neuron 131_1 may generate an output spike signal (an output) based on a comparison result of the comparator 132_1. For example, the neuron 131_1 may output an output spike signal when the voltage Vm of the membrane signal becomes greater (higher) than the voltage Vth of the threshold signal or when the voltage Vm of the membrane signal reaches the voltage Vth of the threshold signal (fire). In another example, the neuron 131_1 may output an output spike signal when the voltage Vm of the membrane signal becomes smaller (lower) than the voltage Vth of the threshold signal or when the voltage Vm of the membrane signal reaches the voltage Vth of the threshold signal (fire).; [0037]
The spike neural network circuit 100_1 may include a capacitor Cm in which charges are accumulated (integrated) by the first to third operation signals (currents) output from the first to third synapses 121_1 to 121_3.; [0038] The spike neural network circuit 100_1 may include a transistor MN1 which discharges the charges accumulated in the capacitor Cm depending on a leakage signal. The transistor MN1 may receive the leakage signal through a gate terminal. The transistor MN1 may be connected between the capacitor Cm and the power supply voltage GND. The transistor MN1 may be connected to the capacitor Cm in parallel. The transistor MN1 may control the rate (speed) at which operation signals output from the first to third synapses 121_1 to 121_3 are accumulated in the capacitor Cm. A voltage of the leakage signal may be pre-defined. The transistor MN1 is illustrated as being an NMOS in FIG. 2, but may be implemented using a PMOS, an NMOS, or a combination of the PMOS and the NMOS.; As shown above, Oh teaches a comparator, which compares the membrane signal of the neuron with the threshold signal by comparing the voltage of the membrane signal with the voltage of the threshold signal. The neuron outputs an output spike signal when the voltage of the membrane signal is greater than the voltage of the threshold signal or when the voltage of the membrane signal reaches the voltage of the threshold signal. Further, Oh teaches a capacitor which accumulates charges and a transistor connected between the capacitor and GND, which discharges the accumulated charges based on the leakage (predefined voltage). This hardware implementation allows for controlling the flow of voltage based on the voltage of the membrane signal and the voltage of the threshold signal.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 17, Querlioz, as modified by Kesari and Oh, teaches The method of claim 1.
Oh teaches wherein the method is implemented in a neuromorphic computing system for pattern recognition applications ([0019] The inventive concept relates to a circuit implemented in a semiconductor device in order to perform an operation of a neural network. A neural network of the inventive concept may be an computing system for pattern recognition applications artificial neural network (ANN) capable of processing data or information in a similar manner wherein the method is implemented in a neuromorphic to a biological neural network. The neural network may include multiple layers including artificial neurons similar to biological neurons, and synapses for connecting the multiple layers.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
Regarding claim 18, Querlioz, as modified by Kesari and Oh, teaches The method of claim 1.
Oh teaches wherein the spiking neural network processes temporal data by maintaining consistent effective thresholds across the plurality of neurons despite manufacturing variations in the hardware ([0054] The capacitor Cq and the transistors MN9 to MN11 and MP9 may configure a quiescence adjustment circuit 134_1 b which lowers the voltage Vm of the membrane signal to the power supply voltage GND. The quiescence adjustment circuit 134_1 b may adjust an interval in which the membrane signal is deactivated or an interval in which the output spike signal is deactivated. A quiescence of the neuron 131_1 may represent a duration (time) during which the voltage Vm of the membrane signal is driven or maintained with the power supply voltage GND corresponding to a reset, or a duration during which the output spike signal is activated and then deactivated. The quiescence may be adjusted based on the second bias signal, the transistor MN10, and a capacity of the capacitor Cq. Even when an input spike signal is activated and operation results are output from the synapses 121 in the quiescence, since the voltage Vm of the membrane signal is maintained as the power supply voltage GND, the operation results may be ignored.; [0059] FIG. 5 exemplarily illustrates a timing diagram showing the operation of a comparator of FIG. 4. FIG. 5 will be described with reference to FIG. 4. In FIG. 5, the horizontal axis represents duration, and the vertical axis may represent either voltage or current.; As shown above, Oh teaches a quiescence adjustment circuit, which adjusts an interval in which the membrane signal is deactivated or an interval in which the output spike signal is deactivated. The quiescence of the neuron may be represent a temporal data duration of time and may be adjusted based on second bias signal, transistor, and a capacity of the capacitor, similar to the maintaining consistent effective thresholds across the plurality of neurons despite manufacturing variations in the hardware of the claimed invention.).
Querlioz, Kesari, and Oh are combinable for the same rationale as set forth above with respect to claim 1.
Conclusion
Applicant's amendment necessitated the new ground(s) of rejection presented in this Office action. Accordingly, THIS ACTION IS MADE FINAL. See MPEP § 706.07(a). Applicant is reminded of the extension of time policy as set forth in 37 CFR 1.136(a).
A shortened statutory period for reply to this final action is set to expire THREE MONTHS from the mailing date of this action. In the event a first reply is filed within TWO MONTHS of the mailing date of this final action and the advisory action is not mailed until after the end of the THREE-MONTH shortened statutory period, then the shortened statutory period will expire on the date the advisory action is mailed, and any extension fee pursuant to 37 CFR 1.136(a) will be calculated from the mailing date of the advisory action. In no event, however, will the statutory period for reply expire later than SIX MONTHS from the date of this final action.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to MAGGIE MAIDO whose telephone number is (703) 756-1953. The examiner can normally be reached M-Th: 6am - 4pm.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Michael Huntley can be reached on (303) 297-4307. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/MM/Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129
Prosecution Insights