DETAILED ACTION
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 .
Continued Examination Under 37 CFR 1.114
A request for continued examination under 37 CFR 1.114, including the fee set forth in 37 CFR 1.17(e), was filed in this application after final rejection. Since this application is eligible for continued examination under 37 CFR 1.114, and the fee set forth in 37 CFR 1.17(e) has been timely paid, the finality of the previous Office action has been withdrawn pursuant to 37 CFR 1.114. Applicant's submission filed on 05/12/2026 has been entered.
Response to Arguments
Applicant's arguments filed 05/12/2026 have been fully considered and they are partially persuasive.
Regarding applicant’s remarks directed to the rejection of claims under 35 USC § 103, the arguments are directed to newly amended limitations that were not previously examined by the examiner. Therefore, applicants arguments are rendered moot. The examiner refers to the rejection under 35 USC § 103 in the current office action for more details.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis (i.e., changing from AIA to pre-AIA ) for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
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.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
Claim(s) 1-2, 5-10, 12-20 are rejected under 35 U.S.C. 103 as being unpatentable over U.S. Pub. No. US20190370653A1 Chakrabartty et al. (“Chakrabartty-2019”) in view of US Pub No. US20220198245A1 Cleland et al. (“Cleland”) and in view of Frenkel, Charlotte, Martin Lefebvre, and David Bol. "Learning without feedback: Fixed random learning signals allow for feedforward training of deep neural networks." Frontiers in neuroscience 15 (2021): 629892. (“Frenkel”) in view of Chatterjee, Oindrila, and Shantanu Chakrabartty. "Resonant machine learning based on complex growth transform dynamical systems." IEEE Transactions on Neural Networks and Learning Systems 32.3 (2020): 1289-1303. (“Chatterjee”)
In regards to claim 1,
Charkrabartty-2019 teaches
A backpropagation-less learning (BPL) computing device comprising at least one processor in communication with a memory device, the at least one processor configured to:
(Chakrabartty-2019, “[0006] In an additional aspect, a spiking SVM is provided that includes a growth transform neural network system. The growth transform neural network system includes a computing device. The computing device includes at least one processor and a memory [computing device comprising at least one processor in communication with a memory device, the at least one processor configured to:] storing a plurality of modules. Each module includes instructions executable on the at least one processor. The plurality of modules includes a growth transform neural network module, a growth transform module, and a network convergence module. The growth transform neural network module defines a plurality of mirrored neuron pairs that include a plurality of first components and a plurality of second components that are interconnected according to an interconnection matrix. The growth transform module updates each first component of each mirrored neuron pair of a plurality of mirrored neuron pairs according to a growth transform neuron model.”)
retrieve, from the memory device, at least one or more training datasets;
(Chakrabartty-2019, “[0142] Processor 404 may also be operatively coupled to a storage device 410 [from the memory device]. Storage device 410 is any computer-operated hardware suitable for storing and/or retrieving data. In some embodiments, storage device 410 is integrated in server computing device 402. For example, server computing device 402 may include one or more hard disk drives as storage device 410. In other embodiments, storage device 410 is external to server computing device 402 and is accessed by a plurality of server computing devices 402. For example, storage device 410 may include multiple storage units such as hard disks or solid state disks in a redundant array of inexpensive disks (RAID) configuration. Storage device 410 may include a storage area network (SAN) and/or a network attached storage (NAS) system.”)
(Chakrabartty-2019, “[0123] While the systems described herein were demonstrated using simple two-dimensional synthetic problems (for the ease of visualization), it is to be understood that the results are suitable for larger and more complex tasks. By way of non-limiting example, Table I summarizes the classification results of different variants of switching SVMs, trained and evaluated on a benchmark ‘Adult (a3a)’ dataset. The training dataset (3185 instances) [at least one or more training datasets] and testing dataset (29376 instances) provided on the LIBSVM website were used for training and cross-validation respectively.”)
build a spike-response model relating one or more aspects of the at least one or more training datasets;
(Charkrabartty-2019, “[0070] In one aspect, the disclosed neural network system is incorporated into a spiking support vector machine (SVM) [build a spike-response model] that includes a network of growth transform neurons, as described herein below. Each neuron in the SVM network learns to encode output parameters such as spike rate and time-to-spike responses according to an equivalent margin of classification and those neurons corresponding to regions near the classification boundary learn to exhibit noise-shaping dynamics similar to behaviors observed in biological networks. As a result, the disclosed spiking support vector machine (SVM) enables large-scale population encoding, for examples for a case when the spiking SVM learns to solve two benchmark classification tasks [relating one or more aspects of the at least one or more training datasets], resulting in classification performance similar to that of the state-of-the-art SVM implementations.”)
store the spike-response model in the memory device;
(Charkrabartty-2019, “[0135] In various aspects, the methods described herein are implemented using a remote and/or local computing device as described herein below. FIG. 21 illustrates an example configuration of a remote device system 300 and depicts an exemplary configuration of a remote or user computing device 302, such as requestor computing device. Computing device 302 includes a processor 304 for executing instructions. In some embodiments, executable instructions are stored in a memory area 306. Processor 304 may include one or more processing units (e.g., in a multi-core configuration). Memory area 306 is any device allowing information such as executable instructions and/or other data to be stored and retrieved [store the spike-response model in the memory device]. Memory area 306 may include one or more computer-readable media.”)
and design, using the spike-response model, a Growth Transform (GT) neural network including a plurality of ON-OFF neuron pairs
(Charkrabartty-2019, Fig. 1C, “[0067]
In various aspects, a growth transform neuron model is mutually coupled with a network objective function of a machine learning model [using the spike-response model; wherein the machine learning model is the spiking machine learning model (SVM)], as illustrated in FIG. 1C. While each individual neuron traverses a trajectory within the dual optimization space as the system converges on a solution, the overall network traverses a trajectory in an equivalent primal optimization space [on overall network spiking activity]. As a result, the network of growth transform neurons solves classification tasks while producing unique but interpretable neural dynamics, such as noise-shaping, spiking and bursting [design… a Growth Transform (GT) neural network].”)
(Chakrabartty-2019, “[0007] …The growth transform neural network module defines a plurality of mirrored neuron pairs [including a plurality of ON-OFF neuron pairs; see fig. 6A for on/off using the sgn function to create the binary values of on/off] that include a plurality of first components and a plurality of second components that are interconnected according to an interconnection matrix. The growth transform module updates each first component of each mirrored neuron pair of a plurality of mirrored neuron pairs according to a growth transform neuron model. The network convergence module converges the plurality of mirrored neuron pairs to a steady state condition by solving a system objective function subject to at least one normalization constraint. The first component and the second component of each mirrored neuron pair in the steady state condition may each produce a neuron response that includes a steady state value or a limit cycle with ΣΔ modulation according to a user-defined potential function Φ(pik) given by:
W 1∈1 +|p ik−(½−∈1)| for 0≤p ik<½−∈1,
W 1 |p ik−½| for ½−∈1 ≤p ik≤½,
W 2 |p ik−½| for ½<p ik<½+∈2, and
W 2∈2 +|p ik−(½−∈2)| for ½+∈2 <p ik≤1,
in which pik is the response of ith neuron of the plurality of mirrored neuron pairs, k is 1 or 2, W1>1, W2>1 ∈1>0, and ∈2>0….
[0022] FIG. 6A is a schematic illustration showing the relationship of the quantized response Sik=sgn(pik−1/M) to the ΣΔ limit cycles.
PNG
media_image1.png
258
527
media_image1.png
Greyscale
”)
enforce sparsity constraints on overall network spiking activity.
In light of the specification, (“[0070] As shown in FIG. 4A, energy-efficiency in energy-based neuromorphic machine learning, there is a loss function for training and an additional loss for enforcing sparsity. The embodiments set forth herein make the loss for training and loss for enforcing sparsity equal, as shown in FIG. 4B”) Examiner interprets the loss function to enforce sparsity constrains on overall network spiking activity.
(Charkrabartty-2019, “[0073] In one aspect, the growth transform neural network design and analysis includes estimating an equivalent dual optimization function based on the mapping given by Eqn. (1). Each neuron implements a continuous mapping based on a polynomial growth transform update that also dynamically optimizes the cost function [enforce sparsity constraints on overall network spiking activity; wherein the cost function corresponds to primal loss-functions]. Because the growth transform mapping is designed to evolve over a constrained manifold, the neuronal responses and the network are stable. The switching, spiking and bursting dynamics of the neurons emerge by choosing different types of potential functions and hyper-parameters in the dual cost function. The use of this approach is suitable for use in the design of SVMs that exhibit ΣΔ modulation type limit-cycles, spiking behavior and bursting responses.”)
Chakrabartty-2019 teaches the cost function corresponding to primal loss-functions
(Chakrabartty-2019, “[0122] One insight that emerged from the disclosed geometric framework is that, while each individual neuron is optimizing a relatively simple dual cost function, the network as a whole exhibits complex dynamics corresponding to primal loss-functions with hysteresis and discontinuities. In each of the support vector networks that incorporate the disclosed growth transform neural networks as described above, irrespective of the nature of the output (ΣΔ modulation, spiking or busting), the output of the neuron faithfully encodes an equivalent classification margin.”)
However, Charkrabartty-2019 does not explicitly teach using weight adaptation only between neurons in each layer of the GT neural network while keeping inter-layer connections fixed to enable simultaneous and independent weight updates for each layer… wherein the sparsity constraints function as a regularizer such that the GT neural network is configured to (i) learn via few-shot learning and (ii) be locally implemented on hardware of a device located at the edge of a network and operate at the edge of the network in energy and resource-constrained environments based at least in part on using the weight adaptation only between neurons in each layer of the GT neural network while keeping the inter-layer connections fixed, such that the GT neural network implemented on the hardware of the device can execute weight updates at the edge of the network without reliance on global information.
Cleland teaches
using weight adaptation only between neurons in each layer of the GT neural network… based at least in part on using the weight adaptation only between neurons in each layer of the GT neural network
(Cleland, “[0314] Input from each sensor was delivered directly to ET and PG interneurons associated with the column corresponding to that sensor, and the resulting PG cell activity was delivered via graded synaptic inhibition onto all ET cells within all columns in the network. ET cells in turn then synaptically excited their corresponding, cocolumnar MCs, as illustrated in FIG. 11. The approximate outcome of this preprocessor algorithm is as follows: given that x1 ET, x2 ET, x3 ET, . . . , xn ET denote the responses of ET cells to odor inputs (prior to their inhibition by PG cells), and x1 pg, x2 pg, x3 pg, . . . , xn pg denote the analogous responses of PG interneurons to these same inputs, the resulting input to MC somata from ET cells following their PG-mediated lateral inhibition [using weight adaptation only between neurons in each layer of the GT neural network; ie lateral inhibition] will be
PNG
media_image2.png
60
516
media_image2.png
Greyscale
”)
such that the GT neural network is configured to (i) learn via few-shot learning
(Cleland, “[0292] While certain of the present embodiments focus on one-shot learning, the network can also be configured for few-shot learning, in which it gradually adapts to the underlying statistics of training samples.”)
and (ii) be locally implemented on hardware of a device located at the edge of a network and operate at the edge of the network in energy and resource-constrained environments
Examiner interprets the limitation as a TinyML device as an edge device to a network in light of the specification (“[0051] In the exemplary embodiment, GT system 114 a-114 c may be tinyML systems, or networks, that implement machine learning processes. In some embodiments, a tinyML system may include a device that provides low latency, low power consumption, low bandwidth, and privacy. Additionally, a tinyML device, sometimes called an always on device, may be placed on the edge of a network.”)
(Cleland, “[0301] Neuromorphic systems are custom integrated circuits that model biological neural computations, typically with orders of magnitude greater speed and energy efficiency than general-purpose computers. These systems enable the deployment of neural algorithms in edge devices [be locally implemented on hardware of a device located at the edge of a network and operate at the edge of the network], such as chemosensory signal analyzers, in which real-time operation, low power consumption [in energy], environmental robustness, and compact size [and resource-constrained environments] are important operational metrics.”)
Chakrabartty-2019 is considered to be analogous to the claimed invention because they are in the same field of GT neural networks. Cleland is considered to be analogous to the claimed invention because they are in the same field of spiking neural networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Cleland in order to provide few-shot learning to learn robust representations of even corrupt samples, provide neuromorphic systems with greater speed and energy efficiency than general-purpose computers and provide lateral inhibition as contrast enhancement. (Cleland, [0292], “In this configuration, the network learns robust representations even when the training samples themselves are corrupted by impulse noise.”) (Cleland, [0301], “Neuromorphic systems are custom integrated circuits that model biological neural computations, typically with orders of magnitude greater speed and energy efficiency than general-purpose computers.”) (Cleland, [0084], “When implemented on appropriate hardware, SNNs are extremely energy efficient and uniquely scalable to very large problems; consequently, marrying the efficiency of neuromorphic processors with the algorithmic power of deep learning is an important industry goal.”) (Cleland, [0308], “Within this first (glomerular) layer of the MOB, a number of preprocessing computations also are performed, including a high-dimensional form of contrast enhancement and an intricate set of computations mediating a type of global feedback normalization that enables concentration tolerance.”)
Frenkel teaches
while keeping inter-layer connections fixed
(Frenkel, Fig. 1 pg. 3 para. 1, “Figure 1: The proposed direct random target projection algorithm builds on feedback-alignment-based algorithms to tackle the weight transport problem while further releasing update locking. Black arrows indicate the feedforward pathways and orange arrows the feedback pathways. In the k-th layer, the weighted sum of inputs yk−1 is denoted as zk, the bias as bk, the activation function as fk(·) and its derivative as fk(·), with k ∈ [1,K], k ∈ N, and K the number of layers. Trainable forward weight matrices are denoted as Wk and fixed random connectivity matrices as Bk [while keeping inter-layer connections fixed; fixed random connectivity matrices]. The input vector is denoted as x, the target vector as y∗ and the loss function as J(·). The estimated loss gradients for the outputs of the k-th hidden layer, denoted as _yk, are provided for each training algorithm. The layer equations for zk, yk and _zk, defined as the modulatory signals, are provided in the upper left corner, with ⊙ denoting the element-wise multiplication operator. (A) Backpropagation of error (BP) algorithm [4]. (B) Feedback alignment (FA) algorithm [13]. (C) Direct feedback alignment (DFA) algorithm [15]. (D) Proposed direct random target projection (DRTP) algorithm. Adapted from [15] and [24].”)
to enable simultaneous and independent weight updates for each layer
Examiner’s note: Examiner interprets the limitation in light of the specification (“[0009] Some prior art solutions have developed algorithms for training neural networks that overcome one or more constraints of the backpropagation algorithm. One known method, feedback alignment or random backpropagation, eradicates the weight transport problem by using fixed random weights in the feedback path for propagating error gradient information. Research showed that directly propagating the output error or the raw one-hot encoded targets is sufficient to maintain feedback alignment, and, in the case of the latter, also eradicates update locking by allowing simultaneous and independent weight updates at each layer.”)
(Frenkel, Section 1, pg. 2 para. 3, “In order to fully solve both the weight transport and the update locking problems, we propose the direct random tar get projection (DRTP) algorithm (Fig. 1D). Compared to DFA, the targets are used in place of the output error and projected onto the hidden layers. We demonstrate both theoretically and experimentally that, in the framework of classification problems, the error sign information contained in the targets is sufficient to maintain feedback alignment with the loss gradients δzk for the weighted sum of inputs in layer k, denoted as the modulatory signals in the subsequent text, and allows training multi-layer networks, leading to three key advantages. First, DRTP solves the weight transport problem by entirely removing the need for dedicated feedback pathways. Second, layers can be updated independently and without update locking [to enable simultaneous and independent weight updates for each layer; ie without update locking] as a full forward pass is not required, thus reducing memory requirements by releasing the need to buffer inputs and activations of each layer. Third, DRTP is a purely feedforward and low-cost algorithm whose updates rely on layerwise information that is immediately available upon computation of the layer outputs.”)
such that the GT neural network implemented on the hardware of the device can execute weight updates at the edge of the network without reliance on global information.
(Frenkel, Abstract, “Based on these observations, we propose the direct random target projection (DRTP) algorithm and demonstrate that it provides a tradeoff between accuracy and computational cost that is suitable for adaptive edge computing devices [such that the GT neural network implemented on the hardware of the device can execute weight updates at the edge of the network; wherein Frenkel discloses the DRTP algorithm is suitable for edge computing devices].”)
(Frenkel, Section 1 pg. 2 para. 3, “Third, DRTP is a purely feedforward and low-cost algorithm whose updates rely on layerwise [without reliance on global information; ie layerwise feedforward training] information that is immediately available upon computation of the layer yk outputs. Estimating the layerwise loss gradients only requires a label-dependent random vector selection, contrasting with the error locality and synthetic gradients approaches that require the addition of side networks for error or gradient prediction. DRTP even compares favorably to DFA, as the latter still requires a multiplication between the output error and a fixed random matrix.”)
Chakrabartty-2019 is considered to be analogous to the claimed invention because they are in the same field of GT neural networks. Frenkel is considered to be analogous to the claimed invention because they are reasonably pertinent to the problem the inventor faced of the bottlenecks of backpropagation (weight transport problem and update locking). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Frenkel in order to provide a direct random target projection (DRTP) algorithm as a solution to the weight transport and update lock problems wherein DRTP is suitable for edge computing devices (Frenkel, Abstract, “While the backpropagation of error algorithm enables deep neural network training, it implies (i) bidirectional synaptic weight transport and (ii) update locking until the forward and backward passes are completed. Not only do these constraints preclude biological plausibility, but they also hinder the development of low-cost adaptive smart sensors at the edge, as they severely constrain memory accesses and entail buffering overhead. In this work, we show that the one-hot-encoded labels provided in supervised classification problems, denoted as targets, can be viewed as a proxy for the error sign. Therefore, their fixed random projections enable a layerwise feedforward training of the hidden layers, thus solving the weight transport and update locking problems while relaxing the computational and memory requirements. Based on these observations, we propose the direct random target projection (DRTP) algorithm and demonstrate that it provides a tradeoff between accuracy and computational cost that is suitable for adaptive edge computing devices.”)
Chatterjee teaches
wherein the sparsity constraints function as a regularizer
(Chatterjee, Section III., “In this formulation, the active-power dissipation D in (10) acts as a regularization function with β ≥ 0 being a hyperparameter…. Example 1: Consider a single-variable quadratic optimization problem of the form H1(x) = x^2, subject to the constraint
|x| ≤ 1, x ∈ R. Substituting x = |V|^2 − |I|^2, the problem can be mapped (please see Appendix B for more details) into the form equivalent to (10) as
PNG
media_image3.png
67
369
media_image3.png
Greyscale
Fig. 4(a)–(c) plots L1 for different values of β [wherein the sparsity constraints are configured to function as a regularizer; wherein L1 optimization acts as a regularization function with β ≥ 0 being a hyperparameter and thus enable the network to perform (i) and (ii). Examiner refers to Chatterjee, Section V for support in combining the cost function of Charkrabartty-2019 with the regularization function of Chatterjee]. As shown in Fig. 4(a) and as expected for β = 0, the cost function has several minima (or attractors), whereas for β > 0, the minima corresponds to φ = ±π/2, for which the active-power dissipation is zero. Fig. 4(b) and (c) shows that controlling β will control the optimization landscape (without changing the location of the attractors) and will determine the attractor trajectory. This feature has been exploited in Sections IV and V to optimize the active-power dissipation profile during the learning phase.”
PNG
media_image4.png
216
750
media_image4.png
Greyscale
PNG
media_image5.png
477
762
media_image5.png
Greyscale
)
Chakrabartty-2019 is considered to be analogous to the claimed invention because they are in the same field of GT neural networks. Chatterjee is considered to be analogous to the claimed invention because they are in the same field of growth transform networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Chatterjee in order to provide a regularization function to select the trajectory with an optimal active-power dissipation profile (Chatterjee, “Illustration showing that operating in the complex domain allows different possible learning trajectories from an initial state to the final steady state. Regularization with respect to the phase factor could then be used to select the trajectory with an optimal active-power dissipation profile and results in limit cycle oscillations in steady state. The circles indicate the constant magnitude loci.”
PNG
media_image6.png
264
364
media_image6.png
Greyscale
)
In regards to claim 2,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Cleland teaches wherein at least one of the at least one or more training datasets has one or more unique challenges include sensor drift, stimulus concentrations, or both.
(Cleland, [0306], “We demonstrate online learning performance using a publicly available machine olfaction dataset with challenges including relatively small training sets, variable stimulus concentrations, and three years of sensor drift [wherein at least one of the at least one or more training datasets has one or more unique challenges include sensor drift, stimulus concentrations, or both].”)
In regards to claim 5,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Chakrabartty-2019 teaches wherein, to enforce sparsity, the at least one processor is configured to: minimize network-level spiking activity while producing classification accuracy comparable to standard approaches on the at least one or more training datasets.
(Charkrabartty-2019, Table 2, “[0125] The first term in the cost function minimizes a kernel distance between the class labels and the probability variables pi +, pi −, and the second term minimizes [minimize network-level spiking activity while producing classification accuracy comparable to standard approaches on the one or more training datasets; wherein Table 2 teaches the classification accuracies is compared to standard approaches ie a benchmark] a cumulative potential function Ω(.) corresponding to each neuron. The kernel or the interconnection matrix Q is a positive definite matrix such that each of its elements is written as an inner-product in a high-dimensional space as Qij=Ψ(xi)·Ψ(xj) where xi∈RD correspond to the input data vector and Ψ(.) represents a high-dimensional mapping function.”)
In regards to claim 6,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Cleland teaches wherein the GT neural network comprises a neuromorphic tinyML system constrained in one or more of an energy, resources, and network structure.
Examiner interprets the limitation as a TinyML device as an edge device to a network in light of the specification (“[0051] In the exemplary embodiment, GT system 114 a-114 c may be tinyML systems, or networks, that implement machine learning processes. In some embodiments, a tinyML system may include a device that provides low latency, low power consumption, low bandwidth, and privacy. Additionally, a tinyML device, sometimes called an always on device, may be placed on the edge of a network.”)
(Cleland, “[0301] Neuromorphic systems are custom integrated circuits that model biological neural computations, typically with orders of magnitude greater speed and energy efficiency than general-purpose computers. These systems enable the deployment of neural algorithms in edge devices, such as chemosensory signal analyzers, in which real-time operation, low power consumption, environmental robustness, and compact size [neuromorphic tinyML system constrained in one or more of an energy, resources] are important operational metrics.”)
In regards to claim 7,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Cleland teaches wherein at least one of the at least one or more training datasets is a publicly available machine olfaction dataset having one or more unique challenges.
(Cleland, [0306], “We demonstrate online learning performance using a publicly available machine olfaction dataset with challenges including relatively small training sets, variable stimulus concentrations, and three years of sensor drift [wherein at least one of the at least one or more training datasets is a publicly available machine olfaction dataset having one or more unique challenges].”)
In regards to claim 8,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Chakrabartty-2019 teaches wherein the spike-response model is built using machine learning, artificial intelligence, or a combination thereof.
(Chakrabartty-2019, “[0067] In various aspects, a growth transform neuron model is mutually coupled with a network objective function of a machine learning model [wherein the spike-response model is built using machine learning; the machine learning model is the spiking machine learning model (SVM)], as illustrated in FIG. 1C.”)
In regards to claim 9,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Chakrabartty-2019 teaches wherein the spike-response model is built using supervised learning, unsupervised learning, or both.
(Chakrabartty-2019, “[0113] The stimuli used were binary labels assigned to the neurons that determine the network configuration for a given classification problem, and the strength of the stimulus is higher for neurons closer to the classification hyperplane (i.e., for the ‘support vectors’) [wherein the spike-response model is built using supervised learning; wherein a support vector machine is a supervised ML algorithm]. As the margin of separation z from the hyperplane decreased, the spiking rate for a support vector increases and it starts spiking earlier in the convergence process.”)
In regards to claim 10,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 9,
Charkrabartty-2019 teaches wherein minimizing a training error is equivalent to minimizing overall spiking activity across the GT neural network.
(Charkrabartty-2019, “[0101] In various aspects, the growth transform neural network is incorporated into the design of a ΣΔ support vector machine (SVM). In this aspect, the potential function is given by Φ(pik)=pik−½|, as shown in FIG. 4A. The gradient of the function Φ(.) in this aspect has a discontinuity at pik=½, k=(1, 2), ∀i. The corresponding primal loss-function corresponding to a binary classification task is obtained using the geometric approach described by Eqn. (24) and illustrated in FIG. 4A. The primal loss-function [equivalent to minimizing overall spiking activity across the GT neural network] exhibits a piece-wise linear response where the slope of the loss-function changes at classification margins (or errors [minimizing a training error]) that are symmetric about the separating hyperplane. For the piece-wise linear potential function, the response of the ith neuron is given by
S ikΦ′(p ik)=sgn(p ik−0.5) (29)
and represents a binary output that switches between two values +1 and −1.”)
In regards to claim 12,
Chakrabartty-2019 teaches
A neuromorphic tinyML system, comprising: at least one [tinyML] device having hardware including a memory and a processor;
(Chakrabartty-2019, “[0004] In one aspect, a growth transform neural network [growth transform (GT) neural network] system is provided that includes a computing device. The computing device includes at least one processor and a memory storing a plurality of modules. Each module includes instructions executable on the at least one processor [device having a memory and a processor]”)
and a growth transform (GT) neural network including a plurality of ON-OFF neuron pairs
(Chakrabartty-2019, “[0007] …The growth transform neural network module defines a plurality of mirrored neuron pairs [including a plurality of ON-OFF neuron pairs; see fig. 6A for on/off] that include a plurality of first components and a plurality of second components that are interconnected according to an interconnection matrix. The growth transform module updates each first component of each mirrored neuron pair of a plurality of mirrored neuron pairs according to a growth transform neuron model. The network convergence module converges the plurality of mirrored neuron pairs to a steady state condition by solving a system objective function subject to at least one normalization constraint. The first component and the second component of each mirrored neuron pair in the steady state condition may each produce a neuron response that includes a steady state value or a limit cycle with ΣΔ modulation according to a user-defined potential function Φ(pik) given by:
W 1∈1 +|p ik−(½−∈1)| for 0≤p ik<½−∈1,
W 1 |p ik−½| for ½−∈1 ≤p ik≤½,
W 2 |p ik−½| for ½<p ik<½+∈2, and
W 2∈2 +|p ik−(½−∈2)| for ½+∈2 <p ik≤1,
in which pik is the response of ith neuron of the plurality of mirrored neuron pairs, k is 1 or 2, W1>1, W2>1 ∈1>0, and ∈2>0….
[0022] FIG. 6A is a schematic illustration showing the relationship of the quantized response Sik=sgn(pik−1/M) to the ΣΔ limit cycles.
PNG
media_image1.png
258
527
media_image1.png
Greyscale
”)
and enforce sparsity constraints on overall network spiking activity,
In light of the specification, (“[0070] As shown in FIG. 4A, energy-efficiency in energy-based neuromorphic machine learning, there is a loss function for training and an additional loss for enforcing sparsity. The embodiments set forth herein make the loss for training and loss for enforcing sparsity equal, as shown in FIG. 4B”) Examiner interprets the loss function to enforce sparsity constrains on overall network spiking activity.
(Charkrabartty-2019, “[0073] In one aspect, the growth transform neural network design and analysis includes estimating an equivalent dual optimization function based on the mapping given by Eqn. (1). Each neuron implements a continuous mapping based on a polynomial growth transform update that also dynamically optimizes the cost function [enforce sparsity constraints on overall network spiking activity; wherein the cost function corresponds to primal loss-functions]. Because the growth transform mapping is designed to evolve over a constrained manifold, the neuronal responses and the network are stable. The switching, spiking and bursting dynamics of the neurons emerge by choosing different types of potential functions and hyper-parameters in the dual cost function. The use of this approach is suitable for use in the design of SVMs that exhibit ΣΔ modulation type limit-cycles, spiking behavior and bursting responses.”)
simultaneously learn optimal parameters for a task and minimize spiking activity across the GT neural network;
(Charkrabartty-2019, “[0073] In one aspect, the growth transform neural network design and analysis includes estimating an equivalent dual optimization function [simultaneously learn optimal parameters for a task] based on the mapping given by Eqn. (1). Each neuron implements a continuous mapping based on a polynomial growth transform update that also dynamically optimizes the cost function. Because the growth transform mapping is designed to evolve over a constrained manifold, the neuronal responses and the network are stable. The switching, spiking [spiking activity across the GT neural network] and bursting dynamics of the neurons emerge by choosing different types of potential functions and hyper-parameters in the dual cost function. The use of this approach is suitable for use in the design of SVMs that exhibit ΣΔ modulation type limit-cycles, spiking behavior and bursting responses.”)
(Charkrabartty-2019, Table 2, “[0125] The first term in the cost function minimizes a kernel distance between the class labels and the probability variables pi +, pi −, and the second term minimizes [minimize] a cumulative potential function Ω(.) corresponding to each neuron. The kernel or the interconnection matrix Q is a positive definite matrix such that each of its elements is written as an inner-product in a high-dimensional space as Qij=Ψ(xi)·Ψ(xj) where xi∈RD correspond to the input data vector and Ψ(.) represents a high-dimensional mapping function.”)
generate a training dataset based on the learned optimal parameters and spiking activity data on the GT neural network;
(Chakrabartty-2019, “[0075] A continuous-time variant of the growth transform neuron model is described herein and a network of growth transform neurons is used to implement a spiking SVM. The growth transform neuron in the SVM network may learn to encode (rate and time-to-spike response) its output according to an equivalent margin of classification and the neurons corresponding to regions near the classification boundary may learn to exhibit noise-shaping dynamics similar to what has been reported in biological networks. In one aspect, the model of the growth transform neuron is summarized along with its dynamical properties. In another aspect, an SVM formulation is mapped onto a growth transform neural network and different spiking dynamics are demonstrated based on synthetic [a training dataset;] and benchmark datasets.”)
Further, the Chakrabartty-2019 teaches the synthetic dataset to be given by a synthetic classification task
(Chakrabartty-2019, “[0111] FIGS. 9B and 9C show the classification contours for a two-dimensional linear and non-linear synthetic [generate] classification task, respectively. The introduction of the non-linearity in the transition region does not affect the classification boundaries or performance. The resulting spike-trains generated by different neurons (located at different classification margins) are shown in FIG. 10 for the system illustrated in FIG. 9B.”)
(Charkrabartty-2019, “[0070] In one aspect, the disclosed neural network system is incorporated into a spiking support vector machine (SVM) that includes a network of growth transform neurons, as described herein below. Each neuron in the SVM network learns to encode output parameters [based on the learned optimal parameters] such as spike rate and time-to-spike responses according to an equivalent margin of classification and those neurons corresponding to regions near the classification boundary learn to exhibit noise-shaping dynamics similar to behaviors observed in biological networks. As a result, the disclosed spiking support vector machine (SVM) enables large-scale population encoding, for examples for a case when the spiking SVM learns to solve two benchmark classification tasks, resulting in classification performance similar to that of the state-of-the-art SVM implementations.”)
(Chakrabartty-2019, “[0114] FIGS. 12A and 12B show that the firing rate (number of spikes per 1000 iterations) monotonically decreases and the time (iteration count) to the first spike monotonically increases with the margin of separation, respectively. T\Without being limited to any particular theory, the nearer a neuron is to the classification hyperplane, the more its relative contribution in determining the weight vectors and the faster it reaches the transition region around p=½, leading to the spiking response. The spikes thus represent a manifestation of the process of convergence, and are directly related to the learning behavior of the dynamical system through the support vectors (i.e., the vectors that are the most useful in learning the classification boundary) [spiking activity data on the GT neural network].”)
design another GT neural network for [at least one other tinyML] device based on the training dataset
(Chakrabartty-2019, “[0075] A continuous-time variant of the growth transform neuron model is described herein and a network of growth transform neurons is used to implement a spiking SVM. The growth transform neuron in the SVM network may learn to encode (rate and time-to-spike response) its output according to an equivalent margin of classification and the neurons corresponding to regions near the classification boundary may learn to exhibit noise-shaping dynamics similar to what has been reported in biological networks. In one aspect, the model of the growth transform neuron is summarized along with its dynamical properties. In another aspect, an SVM formulation is mapped onto a growth transform neural network and different spiking dynamics are demonstrated [design another GT neural network] based on synthetic [based on the training dataset] and benchmark datasets.”)
However, Chakrabartty-2019 does not explicitly teach tinyML… using weight adaptation only between neurons in each layer of the GT neural network …while keeping inter-layer connections fixed to enable simultaneous and independent weight updates for each layer wherein the sparsity constraints function as a regularizer such that the GT neural network is configured to learn via few-shot learning and be locally implemented on hardware of a device located at the edge of a network and operate at the edge of the network in energy and resource-constrained environments based at least in part on using the weight adaptation only between neurons in each layer of the GT neural network while keeping the inter-layer connections fixed, such that the GT neural network implemented on the hardware of the device can execute weight updates at the edge of the network without reliance on global information;
Cleland teaches
for at least one tinyML device having a memory and a processor… for at least one other tinyML device
Examiner’s note: Examiner interprets TinyML systems (and devices) to be energy and resource-constrained.
(Cleland, pg. 78 line 25-pg. 79 line 2, “Neuromorphic systems are custom integrated circuits that model biological neural computations, typically with orders of magnitude greater speed and energy efficiency than general-purpose computers. These systems enable the deployment of neural algorithms in edge devices [for at least one tinyML device having a memory and a processor… for at least one other tinyML device], such as chemosensory signal analyzers, in which real-time operation, low power consumption [energy-constrained], environmental robustness, and compact size [and resource-constrained] are important operational metrics.”)
using weight adaptation only between neurons in each layer of the GT neural network… based at least in part on using the weight adaptation only between neurons in each layer of the GT neural network
(Cleland, “[0314] Input from each sensor was delivered directly to ET and PG interneurons associated with the column corresponding to that sensor, and the resulting PG cell activity was delivered via graded synaptic inhibition onto all ET cells within all columns in the network. ET cells in turn then synaptically excited their corresponding, cocolumnar MCs, as illustrated in FIG. 11. The approximate outcome of this preprocessor algorithm is as follows: given that x1 ET, x2 ET, x3 ET, . . . , xn ET denote the responses of ET cells to odor inputs (prior to their inhibition by PG cells), and x1 pg, x2 pg, x3 pg, . . . , xn pg denote the analogous responses of PG interneurons to these same inputs, the resulting input to MC somata from ET cells following their PG-mediated lateral inhibition [using weight adaptation only between neurons in each layer of the GT neural network; ie lateral inhibition] will be
PNG
media_image2.png
60
516
media_image2.png
Greyscale
”)
such that the GT neural network is configured to learn via few-shot learning
(Cleland, “[0292] While certain of the present embodiments focus on one-shot learning, the network can also be configured for few-shot learning, in which it gradually adapts to the underlying statistics of training samples.”)
and be locally implemented on hardware of a device located at the edge of a network and operate at the edge of the network in energy and resource-constrained environments
Examiner interprets the limitation as a TinyML device as an edge device to a network in light of the specification (“[0051] In the exemplary embodiment, GT system 114 a-114 c may be tinyML systems, or networks, that implement machine learning processes. In some embodiments, a tinyML system may include a device that provides low latency, low power consumption, low bandwidth, and privacy. Additionally, a tinyML device, sometimes called an always on device, may be placed on the edge of a network.”)
(Cleland, “[0301] Neuromorphic systems are custom integrated circuits that model biological neural computations, typically with orders of magnitude greater speed and energy efficiency than general-purpose computers. These systems enable the deployment of neural algorithms in edge devices [be locally implemented on hardware of a device located at the edge of a network and operate at the edge of the network], such as chemosensory signal analyzers, in which real-time operation, low power consumption [in energy], environmental robustness, and compact size [and resource-constrained environments] are important operational metrics.”)
Chakrabartty-2019 is considered to be analogous to the claimed invention because they are in the same field of GT neural networks. Cleland is considered to be analogous to the claimed invention because they are in the same field of spiking neural networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Cleland in order to provide few-shot learning to learn robust representations of even corrupt samples, provide neuromorphic systems with greater speed and energy efficiency than general-purpose computers and provide lateral inhibition as contrast enhancement. (Cleland, [0292], “In this configuration, the network learns robust representations even when the training samples themselves are corrupted by impulse noise.”) (Cleland, [0301], “Neuromorphic systems are custom integrated circuits that model biological neural computations, typically with orders of magnitude greater speed and energy efficiency than general-purpose computers.”) (Cleland, [0084], “When implemented on appropriate hardware, SNNs are extremely energy efficient and uniquely scalable to very large problems; consequently, marrying the efficiency of neuromorphic processors with the algorithmic power of deep learning is an important industry goal.”) (Cleland, [0308], “Within this first (glomerular) layer of the MOB, a number of preprocessing computations also are performed, including a high-dimensional form of contrast enhancement and an intricate set of computations mediating a type of global feedback normalization that enables concentration tolerance.”)
Frenkel teaches
while keeping inter-layer connections fixed
(Frenkel, Fig. 1 pg. 3 para. 1, “Figure 1: The proposed direct random target projection algorithm builds on feedback-alignment-based algorithms to tackle the weight transport problem while further releasing update locking. Black arrows indicate the feedforward pathways and orange arrows the feedback pathways. In the k-th layer, the weighted sum of inputs yk−1 is denoted as zk, the bias as bk, the activation function as fk(·) and its derivative as fk(·), with k ∈ [1,K], k ∈ N, and K the number of layers. Trainable forward weight matrices are denoted as Wk and fixed random connectivity matrices as Bk [while keeping inter-layer connections fixed; fixed random connectivity matrices]. The input vector is denoted as x, the target vector as y∗ and the loss function as J(·). The estimated loss gradients for the outputs of the k-th hidden layer, denoted as _yk, are provided for each training algorithm. The layer equations for zk, yk and _zk, defined as the modulatory signals, are provided in the upper left corner, with ⊙ denoting the element-wise multiplication operator. (A) Backpropagation of error (BP) algorithm [4]. (B) Feedback alignment (FA) algorithm [13]. (C) Direct feedback alignment (DFA) algorithm [15]. (D) Proposed direct random target projection (DRTP) algorithm. Adapted from [15] and [24].”)
to enable simultaneous and independent weight updates for each layer
Examiner’s note: Examiner interprets the limitation in light of the specification (“[0009] Some prior art solutions have developed algorithms for training neural networks that overcome one or more constraints of the backpropagation algorithm. One known method, feedback alignment or random backpropagation, eradicates the weight transport problem by using fixed random weights in the feedback path for propagating error gradient information. Research showed that directly propagating the output error or the raw one-hot encoded targets is sufficient to maintain feedback alignment, and, in the case of the latter, also eradicates update locking by allowing simultaneous and independent weight updates at each layer.”)
(Frenkel, Section 1, pg. 2 para. 3, “In order to fully solve both the weight transport and the update locking problems, we propose the direct random tar get projection (DRTP) algorithm (Fig. 1D). Compared to DFA, the targets are used in place of the output error and projected onto the hidden layers. We demonstrate both theoretically and experimentally that, in the framework of classification problems, the error sign information contained in the targets is sufficient to maintain feedback alignment with the loss gradients δzk for the weighted sum of inputs in layer k, denoted as the modulatory signals in the subsequent text, and allows training multi-layer networks, leading to three key advantages. First, DRTP solves the weight transport problem by entirely removing the need for dedicated feedback pathways. Second, layers can be updated independently and without update locking [to enable simultaneous and independent weight updates for each layer; ie without update locking] as a full forward pass is not required, thus reducing memory requirements by releasing the need to buffer inputs and activations of each layer. Third, DRTP is a purely feedforward and low-cost algorithm whose updates rely on layerwise information that is immediately available upon computation of the layer outputs.”)
such that the GT neural network implemented on the hardware of the device can execute weight updates at the edge of the network without reliance on global information;
(Frenkel, Abstract, “Based on these observations, we propose the direct random target projection (DRTP) algorithm and demonstrate that it provides a tradeoff between accuracy and computational cost that is suitable for adaptive edge computing devices [such that the GT neural network implemented on the hardware of the device can execute weight updates at the edge of the network; wherein Frenkel discloses the DRTP algorithm is suitable for edge computing devices].”)
(Frenkel, Section 1 pg. 2 para. 3, “Third, DRTP is a purely feedforward and low-cost algorithm whose updates rely on layerwise [without reliance on global information; ie layerwise feedforward training] information that is immediately available upon computation of the layer yk outputs. Estimating the layerwise loss gradients only requires a label-dependent random vector selection, contrasting with the error locality and synthetic gradients approaches that require the addition of side networks for error or gradient prediction. DRTP even compares favorably to DFA, as the latter still requires a multiplication between the output error and a fixed random matrix.”)
Chakrabartty-2019 is considered to be analogous to the claimed invention because they are in the same field of GT neural networks. Frenkel is considered to be analogous to the claimed invention because they are reasonably pertinent to the problem the inventor faced of the bottlenecks of backpropagation (weight transport problem and update locking). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Frenkel in order to provide a direct random target projection (DRTP) algorithm as a solution to the weight transport and update lock problems wherein the DRTP is suitable for edge computing devices (Frenkel, Abstract, “While the backpropagation of error algorithm enables deep neural network training, it implies (i) bidirectional synaptic weight transport and (ii) update locking until the forward and backward passes are completed. Not only do these constraints preclude biological plausibility, but they also hinder the development of low-cost adaptive smart sensors at the edge, as they severely constrain memory accesses and entail buffering overhead. In this work, we show that the one-hot-encoded labels provided in supervised classification problems, denoted as targets, can be viewed as a proxy for the error sign. Therefore, their fixed random projections enable a layerwise feedforward training of the hidden layers, thus solving the weight transport and update locking problems while relaxing the computational and memory requirements. Based on these observations, we propose the direct random target projection (DRTP) algorithm and demonstrate that it provides a tradeoff between accuracy and computational cost that is suitable for adaptive edge computing devices.”)
Chatterjee teaches
wherein the sparsity constraints function as a regularizer
(Chatterjee, Section III., “In this formulation, the active-power dissipation D in (10) acts as a regularization function with β ≥ 0 being a hyperparameter…. Example 1: Consider a single-variable quadratic optimization problem of the form H1(x) = x^2, subject to the constraint
|x| ≤ 1, x ∈ R. Substituting x = |V|^2 − |I|^2, the problem can be mapped (please see Appendix B for more details) into the form equivalent to (10) as
PNG
media_image3.png
67
369
media_image3.png
Greyscale
Fig. 4(a)–(c) plots L1 for different values of β [wherein the sparsity constraints are configured to function as a regularizer; wherein L1 optimization acts as a regularization function with β ≥ 0 being a hyperparameter and thus enable the network to perform (i) and (ii). Examiner refers to Chatterjee, Section V for support in combining the cost function of Charkrabartty-2019 with the regularization function of Chatterjee]. As shown in Fig. 4(a) and as expected for β = 0, the cost function has several minima (or attractors), whereas for β > 0, the minima corresponds to φ = ±π/2, for which the active-power dissipation is zero. Fig. 4(b) and (c) shows that controlling β will control the optimization landscape (without changing the location of the attractors) and will determine the attractor trajectory. This feature has been exploited in Sections IV and V to optimize the active-power dissipation profile during the learning phase.”
PNG
media_image4.png
216
750
media_image4.png
Greyscale
PNG
media_image5.png
477
762
media_image5.png
Greyscale
)
Chakrabartty-2019 is considered to be analogous to the claimed invention because they are in the same field of GT neural networks. Chatterjee is considered to be analogous to the claimed invention because they are in the same field of growth transform networks. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Chatterjee in order to provide a regularization function to select the trajectory with an optimal active-power dissipation profile (Chatterjee, “Illustration showing that operating in the complex domain allows different possible learning trajectories from an initial state to the final steady state. Regularization with respect to the phase factor could then be used to select the trajectory with an optimal active-power dissipation profile and results in limit cycle oscillations in steady state. The circles indicate the constant magnitude loci.”
PNG
media_image6.png
264
364
media_image6.png
Greyscale
)
In regards to claim 13,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The neuromorphic tinyML system of claim 12,
Chakrabartty-2019 teaches wherein the GT neural network is further configured to: store the training dataset on a database communicatively-coupled to the GT neural network.
(Chakrabartty-2019, “[0142] Processor 404 may also be operatively coupled to a storage device 410. Storage device 410 is any computer-operated hardware suitable for storing and/or retrieving data.”)
In regards to claim 14,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The neuromorphic tinyML system of claim 13,
Chakrabartty-2019 teaches wherein the training dataset is used to develop a training data model for designing new tinyML systems, wherein the training data model is created using the training dataset and one or more additional datasets.
(Chakrabartty-2019, “[0075] A continuous-time variant of the growth transform neuron model is described herein and a network of growth transform neurons is used to implement a spiking SVM. The growth transform neuron in the SVM network may learn to encode (rate and time-to-spike response) its output according to an equivalent margin of classification and the neurons corresponding to regions near the classification boundary may learn to exhibit noise-shaping dynamics similar to what has been reported in biological networks. In one aspect, the model of the growth transform neuron is summarized along with its dynamical properties. In another aspect, an SVM formulation is mapped onto a growth transform neural network and different spiking dynamics are demonstrated [training data model is created] based on synthetic [using the training dataset] and benchmark datasets [one or more additional datasets].”)
In regards to claim 15,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The neuromorphic tinyML system of claim 14,
Cleland teaches wherein the one or more additional datasets is a publicly-available dataset.
(Cleland, [0306], “We demonstrate online learning performance using a publicly available machine olfaction dataset with challenges including relatively small training sets, variable stimulus concentrations, and three years of sensor drift [wherein the one or more additional datasets is a publicly-available dataset].”)
In regards to claim 16,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The neuromorphic tinyML system of claim 15,
Cleland teaches wherein the publicly- available dataset is a machine olfaction dataset.
(Cleland, [0306], “We demonstrate online learning performance using a publicly available machine olfaction dataset with challenges including relatively small training sets, variable stimulus concentrations, and three years of sensor drift [wherein the publicly- available dataset is a machine olfaction dataset].”)
Claim 17 is rejected on the same rationale under 35 U.S.C. 103 as claim 16 as they are substantially similar (wherein claim 16 is a dependent of claim 14 and developing a training data model is substantially similar to updating it as it uses the same training data set and no additional step is provided for updating)
In regards to claim 18,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Cleland teaches wherein a number of shots of the few-shot learning is between 1 and 10.
(Cleland, [0336], “one-shot, two-shot, three-shot, up through ten-shot [wherein a number of shots of the few-shot learning is between 1 and 10] in order to measure the utility of additional training. Test data (across all trained odorants and all concentrations in the dataset) were classified with 100.0% accuracy in all cases.”)
In regards to claim 19,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Cleland teaches wherein the GT neural network is configured to be embodied within a tinyML device operated at the edge of the network.
(Cleland, [0301], “Neuromorphic systems are custom integrated circuits that model biological neural computations, typically with orders of magnitude greater speed and energy efficiency than general-purpose computers. These systems enable the deployment of neural algorithms in edge devices [wherein the GT neural network is configured to be embodied within a tinyML device operated at the edge of the network], such as chemosensory signal analyzers, in which real-time operation, low power consumption [in energy], environmental robustness, and compact size [and resource-constrained] are important operational metrics.”)
In regards to claim 20,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 19,
Cleland teaches wherein the tinyML device is configured to perform learning
(Cleland, “The theoretical promise of SNNs has motivated the development of multiple neuromorphic hardware platforms, including the academic platforms SpiNNaker and BrainScaleS, as well as IBM’s TrueNorth platform and, most recently, the Loihi platform of Intel Labs, which embeds rapid, iterative plasticity (“self-learning”) [wherein the tinyML device is configured to perform learning] in a compact, scalable chip. These hardware platforms are considered illustrative examples of what are more generally referred to herein as“processing devices” or“processing platforms.” A premise of these hardware projects is that the availability of these platforms would spur the development of practical, useful neuromorphic algorithms by highlighting their particular strengths in energy efficiency and scalability.”)
Cleland teaches and inference at the edge of the network.
(Cleland, “Completing one inference cycle [and inference] (sniff; 5 gamma cycles; 200 timesteps) of the 72-core network required 2.75 ms and consumed 0.43 mJ, of which 0.12 mJ is dynamic energy. It should be noted that the time required to solution was not significantly affected by the scale of the problem (FIG. 10(f)), owing to the Loihi architecture’s fine-grained parallelism. This scalability highlights a key advantage of neuromorphic hardware for application to computational neuroscience and machine olfaction. Energy consumption also scaled only modestly as network size increased (FIG. 10(g)), owing to the colocalization of memory and compute and the use of sparse (spiking) communication, which minimize the movement of data. Using multichip Loihi systems [at the edge of the network], illustrative embodiments are readily scalable to hundreds of columns and hundreds of thousands of intemeurons, and can integrate circuit models of the glomerular layer and the piriform cortex with the current EPL network of the MOB.”)
Claim(s) 3-4 are rejected under 35 U.S.C. 103 as being unpatentable over Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee in further view of O. Chatterjee and S. Chakrabartty, "Decentralized Global Optimization Based on a Growth Transform Dynamical System Model," (“Chakrabartty-2018”).
In regards to claim 3,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Chakrabartty-2018 teaches wherein the spike-response model further includes a learning framework comprising: spike responses generated
(Chakrabartty-2018, Section II., “Consider a functional optimization problem of the following form:
PNG
media_image7.png
128
504
media_image7.png
Greyscale
where q:RM↦R is any arbitrary function having multiple local minima, but a single global minimum at x∗ (i.e., q(x∗)<q(x),∀x∈D ), and H:RM+↦R is a convex functional with respect to f(x) . Typically, we choose H{q(x),f(x)} as some distance measure between the functions q(x) and f(x) , e.g., H{q(x),f(x)}=∫x∈D{q(x)−f(x)}2dx,∀x∈D . Note that (5) involves optimizing a convex cost functional H over a convex domain and thus has a unique solution. The Lagrangian L1 [spike responses generated] of (5) is given by
PNG
media_image8.png
75
505
media_image8.png
Greyscale
”)
Chakrabartty-2018 teaches as a result of a constraint violation;
(Chakrabartty-2018, Section II., “Taking derivative with respect to f(x) on both sides and making use of the fact that H is convex [implying that its derivative M{q,f,β} is monotonic with respect to f(x) ], we have
PNG
media_image9.png
35
358
media_image9.png
Greyscale
where [Φ{⋅,⋅}]+=max(0,Φ{⋅,⋅}) denotes the hinge function. Φ{⋅,⋅} is monotonically decreasing in q , implying Φ{q(x∗),β}>Φ{q(x),β} for q(x)>q(x∗)∀x≠x∗ . Thus, from the constraint equation [as a result of a constraint violation] ∫xf(x)dx=ν , we have
PNG
media_image10.png
59
369
media_image10.png
Greyscale
”)
Chakrabartty-2018 teaches one or more optimal parameters for a certain task learned using neurally relevant local learning rules;
(Chakrabartty-2018, Fig. 1, Illustration of different annealing principles for optimizing a target objective function q(x) with multiple local minima [using neurally relevant local learning rules] but only one global minimum x∗ . (a) Process of surmounting the energy barriers in simulated annealing. (b) Process of quantum tunneling through the barriers in quantum annealing. (c) Proposed approach where a driver function h(x,t) evolves under the influence of q(x) [one or more optimal parameters for a certain task learned; wherein the optimal parameters are the parameters of q(x)] to a Dirac-delta function centered at the global minimum x∗ .”)
PNG
media_image11.png
348
504
media_image11.png
Greyscale
However, Chakrabartty-2018 does not explicitly teach network optimization to encode a solution with as few spikes as possible; and a framework that is operable to incorporate additional structural and connectivity constraints on the GT neural network
Chakrabartty-2019 teaches network optimization to encode a solution with as few spikes as possible; and a framework that is operable to incorporate additional structural and connectivity constraints on the GT neural network.
(Chakrabartty-2019, “[0084]
This formulation is now consistent with the multi-class probability regression framework which was used for deriving different variants of SVMs. Introducing variables yjk, Eqn. (10) is expressed as:
Ψ−1(p ik)=Σj Q ij(p jk +y jk) (12)
where bik satisfies the relation:
b ik=−Σj Q ij y jk ,k=1,2 (13)
under the assumption that Q−1 exists. Eqn. (12) along with the constraints [a framework that is operable to incorporate additional structural and connectivity constraints on the GT neural network] given by Eqns. (5) and (6) are viewed as a first-order condition for the following minimization problem:
PNG
media_image12.png
71
304
media_image12.png
Greyscale
where
Φ(p ik)=−∫Ψ−1(p ik)dp ik. (16)
[0085]Φ(.) is referred to as the potential function. If it is assumed that the matrix Q is positive-definite, the first part of the optimization function in Eqn. (15) is equivalent to minimizing a quadratic distance between the responses pik and the variables yik. The second part of the optimization function is equivalent to minimizing a cumulative potential function Φ(.) corresponding to each neuron [network optimization to encode a solution with as few spikes as possible].”)
Chakrabartty-2018 is considered to be analogous to the claimed invention because they are in the same field of growth transform models. Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Chakrabartty-2018 in order to provide a dynamical system that exploits constraints and solve discrete global optimization problems in neural networks (Chakrabartty-2018, Abstract, “Conservation principles, such as conservation of charge, energy, or mass, provide a natural way to couple and constrain spatially separated variables. In this paper, we propose a dynamical system model that exploits these constraints for solving nonconvex and discrete global optimization problems. Unlike the traditional simulated annealing or quantum annealing-based global optimization techniques, the proposed method optimizes a target objective function by continuously evolving a driver functional over a conservation manifold, using a generalized variant of growth transformations. As a result, the driver functional asymptotically converges toward a Dirac-delta function that is centered at the global optimum of the target objective function. In this paper, we provide an outline of the proof of convergence for the dynamical system model and investigate different properties of the model using a benchmark nonlinear optimization problem. Also, we demonstrate how a discrete variant of the proposed dynamical system can be used for implementing decentralized optimization algorithms, where an ensemble of spatially separated entities (for example, biological cells or simple computational units) can collectively implement specific functions, such as winner-take-all and ranking, by exchanging signals only with its immediate substrate or environment. The proposed dynamical system model could potentially be used to implement continuous-time optimizers, annealers, and neural networks.”)
In regards to claim 4,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee and Chakrabartty-2018 teaches The BPL computing device of claim 3,
Chakrabartty-2018 teaches wherein the spike responses are Lagrangian parameters.
(Chakrabartty-2018, Section II., “Consider a functional optimization problem of the following form:
PNG
media_image7.png
128
504
media_image7.png
Greyscale
where q:RM↦R is any arbitrary function having multiple local minima, but a single global minimum at x∗ (i.e., q(x∗)<q(x),∀x∈D ), and H:RM+↦R is a convex functional with respect to f(x) . Typically, we choose H{q(x),f(x)} as some distance measure between the functions q(x) and f(x) , e.g., H{q(x),f(x)}=∫x∈D{q(x)−f(x)}2dx,∀x∈D . Note that (5) involves optimizing a convex cost functional H over a convex domain and thus has a unique solution. The Lagrangian L1 [Lagrangian parameters] of (5) is given by
PNG
media_image8.png
75
505
media_image8.png
Greyscale
”)
Claim(s) 11 is rejected under 35 U.S.C. 103 as being unpatentable over Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee in further view of S. Farshchi, A. Pesterev, W. -L. Ho and J. W. Judy, "Acquiring High-Rate Neural Spike Data with Hardware-Constrained Embedded Sensors," (“Farshchi”).
In regards to claim 11,
Chakrabartty-2019 in view of Cleland, Frenkel and Chatterjee teaches The BPL computing device of claim 1,
Chakrabartty-2019 teaches further comprising one or more [miniaturized] sensors and devices.
(Chakrabartty-2019, “[0004] In one aspect, a growth transform neural network [growth transform (GT) neural network] system is provided that includes a computing device. The computing device includes at least one processor and a memory storing a plurality of modules. Each module includes instructions executable on the at least one processor [device having a memory and a processor]”)
(Chakrabartty-2019, “[0136] Computing device 302 also includes at least one media output component 308 for presenting information to a user 310….
[0137] In some embodiments, client computing device 302 [this is the same as computing device 302] includes an input device 312 for receiving input from user 310. Input device 312 may include, for example, a keyboard, a pointing device, a mouse, a stylus, a touch sensitive panel (e.g., a touch pad or a touch screen), a camera, a gyroscope, an accelerometer, a position detector, and/or an audio input device [wherein the GT neural network includes one or more [miniaturized] sensors and devices].” Wherein an accelerometer is a sensor for speed, etc)
However, Chakrabartty-2019 does not explicitly teach miniaturized
In light of the specification, (“[0004] Deployment of miniaturized and battery-powered sensors and devices has become ubiquitous and computation is increasingly moving from the cloud to the source of data collection. With it, there is a growing demand for specialized algorithms, hardware and software, collectively termed as tinyML systems. TinyML systems typically can perform learning and inference at the edge in energy and resource- constrained environments. Prior efforts at reducing energy requirements of classic machine learning algorithms include network architecture search, model compression through energy-aware pruning and quantization, model partitioning, among others.”),
Examiner is interpreting miniaturized sensors and devices to be any sensor and device that is energy and resource-constrained.
Farshchi teaches miniaturized sensors and devices
(Farshchi, Section II B. TinyOS and the Mica-Based Sensor Network, “A basic wireless neural recording, archiving, and hosting system based on embedded sensors has been demonstrated; however, continuous signal transmission limits battery life and spike recording to a single channel [5]. A similar vital sign monitoring system that uses TelosB motes has also been demonstrated in [19]; however, the fastest signal it acquires is a single channel of ECG. The work in this paper has been directed toward investigating computationally-efficient software filters and compression algorithms to enable chronic, multi-channel wireless biosignal recording with embedded sensors [sensors and devices] that are hardware and bandwidth constrained [miniaturized].”)
Farshchi is considered to be analogous to the claimed invention because Farshchi is reasonably pertinent to the problem the inventor faced (recording signals representing human functions on resource-constrained hardware). See MPEP 2141.01(a)(i). Therefore, it would have been obvious to someone of ordinary skill in the art before the effective filing date of the claimed invention to have modified Chakrabartty-2019 to incorporate the teachings of Farshchi in order to provide embedded sensors to record data for training a neural network in a way to alleviate network traffic and increase battery life (Farshchi, Abstract, “In an effort to enable embedded sensors that are hardware and bandwidth constrained to acquire high- frequency neural signals, signal-filtering and signal- compression algorithms have been implemented and tested on a commercial-off-the-shelf embedded-system platform. The sensor modules have been programmed to acquire, filter, and transmit raw biological signals at a rate of 32 kbps. Furthermore, on-board signal processing enables one channel sampled at a rate of 4 kS/s at 12-bit resolution to be compressed via ADPCM and transmitted in real time. In addition, the sensors can be configured to only transmit individual time-referenced "spike" waveforms, or only the spike parameters for alleviating network traffic and increasing battery life.”)
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
NPL: Crafton, Brian, et al. "Direct feedback alignment with sparse connections for local learning." Frontiers in neuroscience 13 (2019): 525.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to JASMINE THAI whose telephone number is (703)756-5904. The examiner can normally be reached M-F 8-4.
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 at (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.
/J.T.T./Examiner, Art Unit 2129
/MICHAEL J HUNTLEY/Supervisory Patent Examiner, Art Unit 2129
Prosecution Insights