SRI InternationalDownload PDFPatent Trials and Appeals BoardSep 30, 20212021002555 (P.T.A.B. Sep. 30, 2021) Copy Citation UNITED STATES PATENT AND TRADEMARK OFFICE UNITED STATES DEPARTMENT OF COMMERCE United States Patent and Trademark Office Address: COMMISSIONER FOR PATENTS P.O. Box 1450 Alexandria, Virginia 22313-1450 www.uspto.gov APPLICATION NO. FILING DATE FIRST NAMED INVENTOR ATTORNEY DOCKET NO. CONFIRMATION NO. 15/625,578 06/16/2017 Sek M. Chai SRI7429-2 1038 14824 7590 09/30/2021 Moser Taboada/ SRI International 1030 Broad Street Suite 203 Shrewsbury, NJ 07702 EXAMINER KUO, CHARLES C ART UNIT PAPER NUMBER 2126 NOTIFICATION DATE DELIVERY MODE 09/30/2021 ELECTRONIC Please find below and/or attached an Office communication concerning this application or proceeding. The time period for reply, if any, is set in the attached communication. Notice of the Office communication was sent electronically on above-indicated "Notification Date" to the following e-mail address(es): docketing@mtiplaw.com llinardakis@mtiplaw.com PTOL-90A (Rev. 04/07) UNITED STATES PATENT AND TRADEMARK OFFICE BEFORE THE PATENT TRIAL AND APPEAL BOARD Ex parte SEK M. CHAI, DAVID C. ZHANG, MOHAMED R. AMER, TIMOTHY J. SHIELDS, ASWIN NADAMUNI RAGHAVAN, and BHASKAR RAMAMURTHY Appeal 2021-002555 Application 15/625,578 Technology Center 2100 Before JAMES R. HUGHES, STACEY G. WHITE, and JOHN P. PINKERTON, Administrative Patent Judges. HUGHES, Administrative Patent Judge. DECISION ON APPEAL Appeal 2021-002555 Application 15/625,578 2 STATEMENT OF THE CASE Claims 1–29 are pending1, stand rejected, are appealed by Appellant, and are the subject of our decision under 35 U.S.C. § 134(a).2 See Final Act. 1–2.3 We have jurisdiction under 35 U.S.C. § 6(b). We AFFIRM-IN-PART. CLAIMED SUBJECT MATTER The claimed subject matter, according to Appellant, “relates generally to computer processing systems, and more particularly, to the use of one or more deep neural networks to control and/or inform computer microarchitecture or otherwise manage computing operations.” Spec. ¶ 3. Specifically, Appellant’s claims are directed to systems, computer- readable media, and methods for optimizing operations of computing devices. The method receives a first set of deep neural network inputs (for a first set of deep neural networks (DNNs)), which include: sensor data related to the operation of the computing devices, processing data corresponding to processor instructions, and/or DNN output data. The method generates a 1 The Examiner references claim 29 as having been canceled. Final Act. 1– 2. Claim 29 is pending and appealed. See Appeal Br. 5; Appellant’s Amendment filed April 16, 2020 at page 9. We correct the Examiner’s listing of the claims for clarity and consistency of the record. 2 We use the term “Appellant” to refer to “applicant” as defined in 37 C.F.R. § 1.42. Appellant identifies the real party in interest as SRI International. Appeal Br. 1. 3 We refer to Appellant’s Specification (“Spec.”), filed June 16, 2017 (claiming benefit of US 62/351,205, filed June 16, 2016); Appeal Brief (“Appeal Br.”), filed Sept. 8, 2020; and Reply Brief (“Reply Br.”), filed Mar. 1, 2021. We also refer to the Examiner’s Final Office Action (“Final Act.”), mailed Feb. 21, 2020; and Answer (“Ans.”) mailed Dec. 31, 2020. Appeal 2021-002555 Application 15/625,578 3 first set of DNN outputs using one or more DNNs and the first set of inputs. The DNNs (used to generate the outputs) include DNN parameters learned using one or more computing workloads, and the output generation step analyzes one or more relationships between the received inputs to generate the outputs. The DNN outputs are then provided as: control signals for managing the computing devices, predictions used to generate control signals for managing the computing devices, warnings (indicating that a predicted state a component of the computing device(s) deviates from an acceptable state), and/or a second set of inputs to the first set of DNNs and/or to a second set of DNNs. Spec. ¶¶ 6–45; Abstr. Claim 1 (directed to a method), claim 17 (directed to a system), claim 28 (directed to a computer- readable medium), and claim 29 (directed to a method) are independent. Claim 1, reproduced below, is illustrative of the claimed subject matter: 1. A computer-implemented method for optimizing operations of one or more computing devices having one or more processing units, the method comprising: receiving, as inputs to a first set of one or more deep neural networks (DNNs), at least one of computing environment data selected from a group consisting of: 1) sensor data corresponding with operation of one or more of the computing devices, 2) processing data corresponding with a set of one or more instructions received at one or more of the processing units for execution by the one or more processing units, 3) DNN data corresponding with DNN outputs generated by one or more DNNs involved in managing operations of the one or more computing devices; and 4) any combination of 1), 2), and 3); Appeal 2021-002555 Application 15/625,578 4 applying one or more DNNs to the first set of inputs to generate a first set of one or more DNN outputs based on one or more relationships between the received inputs, the one or more DNNs including DNN parameters learned using one or more computing workloads; and providing the first set of DNN outputs, generated by application of the one or more DNNs, as one or more signals, at least one signal selected from a group consisting of: i) control signals for managing one or more operations of one or more of the computing devices; ii) predictions for use in generating control signals for managing one or more operations of the computing devices, the predictions corresponding to a future state of one or more of the computing devices; iii) warnings indicating that a predicted future state of one or more components of the one or more computing devices deviates more than a threshold deviation level from an acceptable state; iv) a second set of one or more inputs to the first set of DNNs and/or to a second set of one or more DNNs involved in managing operations of the one or more computing devices for generation of a second set of one or more DNN outputs; v) any combination of i), ii), iii), and iv); whereby the one or more signals are provided to the one or more computing devices to enhance performance, efficiency, and/or security of one or more of the computing devices. Appeal Br. 24 (Claims App.) (emphasis added). Appeal 2021-002555 Application 15/625,578 5 REFERENCES The prior art relied upon by the Examiner is: Name Reference Date Bogojeska et al. (“Bogojeska”) US 2015/0178637 A1 June 25, 2015 W. Chedid and Y. Chansu, Survey on Power Management Techniques for Energy Efficient Computer Systems, Cleveland State University (2002) (“Chedid”). C. Hughes and T. Li, Accelerating Multi-core Processor Design Space Evaluation Using Automatic Multi-threaded Workload Synthesis, IEEE Int’l Symposium on Workload Characterization, 163–172, (2008) (“Hughes”). R. Salakhutdinov and G. Hinton, Deep Boltzmann Machines, Proceedings of the Twelfth Int’l Conf. on Artificial Intelligence and Statistics (AISTATS), Journal of Machine Learning Research (JMLR) 5:448-455, 2009 (“Salakhutdinov”) Sheikh et al., Transients of a Micro-Grid System with Multiple Distributed Energy Resources, Sustainable Computing: Informatics and Systems Vol. 2, Issue 3, 151169 (Sept. 2012) (“Sheikh”). REJECTIONS4 1. The Examiner rejected claims 1, 2, 4, 9–15, 17, 18, 21–25, and 27–29 under 35 U.S.C. § 103 as being unpatentable over Bogojeska and Salakhutdinov. See Final Act. 8–33. 2. The Examiner rejected claims 3 and 5 under 35 U.S.C. § 103 as being unpatentable over Bogojeska, Salakhutdinov, and Hughes. See Final Act. 33–35. 4 The Leahy-Smith America Invents Act (“AIA”), Pub. L. No. 112–29, 125 Stat. 284 (2011), amended 35 U.S.C. § 103. Because the present application has an effective filing date after the AIA’s effective date (March 16, 2013), this decision refers to 35 U.S.C. § 103. Appeal 2021-002555 Application 15/625,578 6 3. The Examiner rejected claims 6–8, 19, and 20 under 35 U.S.C. § 103 as being unpatentable over Bogojeska, Salakhutdinov, and Chedid. See Final Act. 35–39. 4. The Examiner rejected claims 16 and 26 under 35 U.S.C. § 103 as being unpatentable over Bogojeska, Salakhutdinov, and Sheikh. See Final Act. 39–41. ANALYSIS Obviousness Appellant argues independent claims 1, 17, 28, and 29 together as a group and does not separately argue the independent claims with particularity. See Appeal Br. 5–15. We select, unless explicitly noted, independent claim 1 as representative of Appellant’s arguments with respect to independent claims 1, 17, 28, and 29. See 37 C.F.R. § 41.37(c)(1)(iv). Appellant makes substantive arguments with respect to several dependent claims. See Appeal Br. 15–23. We address Appellant’s substantive arguments seriatim, unless otherwise noted. Claims 1, 17, 28, and 29 The Examiner rejects independent claim 1 as being obvious in view of Bogojeska and Salakhutdinov. See Final Act. 8–14; Ans. 3–12. Specifically, the Examiner relies on Bogojeska to teach the recited features of claim 1, except for explicitly teaching the disputed deep neural networks (DNNs). See Final Act. 9–14. The Examiner relies on Salakhutdinov to teach the DNNs. See Final Act. 14. In particular, the Examiner relies on Bogojeska to teach a processor, processing unit (claim 1), and hardware processing unit (claim 17); predictive modeler including a neural network; neural network inputs including computing environment data; learned neural Appeal 2021-002555 Application 15/625,578 7 network parameters; and generated neural network outputs. See Final Act. 9–14; Ans. 3–12. Appellant contends Bogojeska and Salakhutdinov do not teach certain disputed features of claim 1 (and claim 17). Appeal Br. 5–15; Reply Br. 2– 11. Specifically, Appellant contends: i) Bogojeska does not teach a processor (processing unit or hardware processor (claim 17)) because “the broadest reasonable interpretation of ‘processor’ or ‘processing unit’ as recited in the claims does not encompass a network of servers” (Reply Br. 2; see Appeal Br. 5–7, 11–12; Reply Br. 3–4); ii) Bogojeska does not teach DNNs that generate DNN outputs based on relationships between the received inputs (see Appeal Br. 7–10; Reply Br. 5–7); iii) Bogojeska does not teach learned DNN parameters (see Appeal Br. 10–11; Reply Br. 7–9); iv) Bogojeska does not teach a DNN configured to receive DNN inputs (see Appeal Br. 12–13; Reply Br. 9–10); and v) Salakhutdinov is not properly combined with Bogojeska (see Appeal Br. 13–15; Reply Br. 10–11). We adopt as our own (1) the findings and reasons set forth by the Examiner in the action from which this appeal is taken (Final Act. 8–14) and (2) the reasons set forth by the Examiner in the Examiner’s Answer (Ans. 3– 12) in response to Appellant’s Appeal Brief. We concur with the findings and conclusions reached by the Examiner, and we provide the following analysis for emphasis. We begin by noting that Appellant does not dispute the Examiner’s findings with respect to Salakhutdinov. See Appeal Br. 5–15; Reply Br. 2– 11. Appellant’s contentions instead focus on the combinability of Salakhutdinov with Bogojeska or the Examiner’s findings with respect to Appeal 2021-002555 Application 15/625,578 8 Bogojeska (supra). We address each of Appellant’s contentions (supra), as follows. i) Processing Unit Appellant’s arguments are not commensurate with the scope of the claims (e.g., claim 1). Even if we accept Appellant’s interpretation of Bogojeska—that “Bogojeska is concerned with ‘modifying server configurations’ in a system that includes a plurality of servers” (Appeal Br 7 (quoting Bogojeska ¶ 2); see Appeal Br. 5–7, 11–12; Reply Br. 3–4)— Bogojeska explicitly discloses a computer system including a processor (see Bogojeska, Fig. 1 (elements 100, 105); Bogojeska ¶¶ 14–15). Appellant’s claims do not preclude the recited “one or more computing devices having one or more processing units” (claim 1) (Appeal Br. 24 (Claims App.)) from including a server or multiple servers. ii) Deep Neural Networks (DNNs) and Generating Outputs Based on Input Relationships Appellant contends Bogojeska does not teach DNNs that generate DNN outputs based on relationships between the received inputs. See Appeal Br. 7–10; Reply Br. 5–7. We disagree. Appellant does not address the Examiner’s findings with respect to Salakhutdinov (supra). As explained by the Examiner, Salakhutdinov describes deep Boltzmann machines, i.e., deep neural networks. See Appeal Br. 14; Ans. 4; Salakhutdinov, Abstr. (p. 1 (5:448)) and Salakhutdinov Sec. 4.1 (p. 6 (5:553)). As further explained by the Examiner, Bogojeska describes a neural network—i.e., a predictive modeler that includes a neural network. See Final Act. 10; Ans. 7–8 (citing Bogojeska ¶ 31). The Examiner relied on the combination of Bogojeska and Salakhutdinov for the teaching of DNNs Appeal 2021-002555 Application 15/625,578 9 (supra). We address Appellant’s combinability arguments in greater detail, infra, but Appellant’s contentions do not persuade us of Examiner error with respect to the DNNs. With respect to the generating outputs based on input relationships, the Examiner finds that Bogojeska describes the operation of its predictive modeler (including a neural network) generating outputs based on inputs that have a relationship and, accordingly, that the outputs are generated based on a relationship between the inputs. See Final Act. 10–12; Ans. 6–8; Bogojeska ¶¶ 32, 36–37. As explained by the Examiner (id.), Bogojeska’s predictive modeler can be trained to predict the quantity and severity of incidents related to a particular feature (such as feature extracted from a hardware specification (see Bogojeska ¶¶ 30)) using historical data related to the feature (e.g., hardware performance data). See Bogojeska ¶ 32. Bogojeska’s predictive modeler then predicts whether a feature is potentially problematic (or unproblematic) and how it affects future operation. Bogojeska ¶¶ 32, 36–37. That is, Bogojeska describes how changes in certain features (e.g., features related to hardware configurations) affect operation of the system. Thus, we find Bogojeska would have at least suggested generating outputs based on input relationships. Appellant’s contentions do not persuade us of Examiner error with respect to generating outputs based on input relationships. iii) DNN Parameters Appellant also contends Bogojeska does not teach learned DNN parameters because the recited DNN parameters are learned using computing workloads, which Appellant contends must be “task[s] . . . in which the processor is engaged” (Appeal Br. 10), and the structure of the Appeal 2021-002555 Application 15/625,578 10 neural network in Bogojeska’s predictive modeler is a “black box” that is not described by Bogojeska (see Appeal Br. 10–11). See Appeal Br. 10–11; Reply Br. 7–9. Contrary to Appellant’s contentions, and as explained by the Examiner (see Final Act. 11–12; Ans. 8–10), Bogojeska describes training the predictive modeler (including the neural network) using historical data, such as hardware performance data to predict future operation of the system. Bogojeska ¶¶ 32, 36–37. Accordingly, Bogojeska describes training the predictive modeler (neural network) to predict operation of the system for certain inputs related to hardware performance. Although Bogojeska does not explicitly describe a computing workload and learned parameters, we find Bogojeska at least suggests that training the predictive modeler would cause “parameters” to be learned and that the learning (learned parameters) would be used in the predicting functions. Further, Salakhutdinov describes training DNNs (learning algorithms for DNNs) at length. See Salakhutdinov Abstr., Salakhutdinov et seq. Thus, the combination of Bogojeska and Salakhutdinov would have at least suggested learned DNN parameters. Further, Appellant’s claim 1 (as well as claim 28) do not require the training information (computing workload) to be processor tasks, or specific processor activity. We agree with the Examiner that Bogojeska’s historical training data may include software and hardware performance data, and that a computing workload can be broadly, but reasonably interpreted to include such historical performance data. See Final Act. 11–12; Ans. 8–9; Bogojeska ¶¶ 32, 36–37. To the extent Appellant contends that Bogojeska does not teach “DNN parameters corresponding to processor behavior” (claim 28) (Appeal Br. 29 (Claims App.)), as discussed, supra, Bogojeska describes how Appeal 2021-002555 Application 15/625,578 11 changes in certain features related to hardware configurations affect operation of the system. Thus, Bogojeska (in combination with Salakhutdinov) would have at least suggested learning DNN parameters that correspond to hardware (e.g., processor) behavior. Appellant’s contentions do not persuade us of Examiner error with respect to learned DNN parameters. iv) Received DNN Inputs Appellant further contends Bogojeska does not teach a DNN configured to receive DNN inputs because, according to Appellant, Bogojeska’s “‘Inputs’ that have been standardized, stored in a database, and fed through a feature extractor are not inputs that are received by a DNN.” Appeal Br. 12; see Appeal Br. 12–13; Reply Br. 9–10. As explained by the Examiner (supra), Bogojeska (in combination with Salakhutdinov) describes receiving inputs to the predictive modeler (neural network) related to hardware performance. See Final Act. 10–12; Ans. 6–8, 11; Bogojeska ¶¶ 32, 36–37. The computing environment data (including, e.g., sensor data) recited in the claims can be broadly, but reasonably interpreted to include such hardware performance data. Additionally, Salakhutdinov describes the operation of, and inputs to, DNNs in detail. See Salakhutdinov Abstr., Salakhutdinov et seq. Thus, Bogojeska (in combination with Salakhutdinov) would have at least suggested receiving DNN inputs that correspond to the recited computing environment data. Appellant’s contentions do not persuade us of Examiner error with respect to a DNN receiving DNN inputs. iv) Combinability Finally, Appellant contends the Examiner did not properly combine Bogojeska and Salakhutdinov. See Appeal Br. 13–15; Reply Br. 10–11. Appeal 2021-002555 Application 15/625,578 12 Appellant’s generalized contentions that Salakhutdinov is concerned with a different purpose or problem (“improves the learning procedure,” which “deals with model creation, not model use” (Appeal Br. 13)), and that the Examiner has not “establish[ed] that the cited combination would lead to predictable results with a reasonable expectation of success” (Appeal Br. 14) do not persuade us that the Examiner erred. See Appeal Br. 13–15; Reply Br. 10–11. The Examiner cited Salakhutdinov to show that deep neural networks and their operation were known. We agree with the Examiner that it would have been within the skill of one of ordinary skill in the art to combine such known techniques of using a deep neural network, as taught by Salakhutdinov, with the system and predictive modeler (neural network) taught by Bogojeska. See KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398, 417 (2007) (“[I]f a technique has been used to improve one device, and a person of ordinary skill in the art would recognize that it would improve similar devices in the same way, using the technique is obvious unless its actual application is beyond his or her skill”). Appellant has not presented evidence sufficient to persuade us that combining the respective familiar elements of the cited references in the manner proffered by the Examiner would have been “uniquely challenging or difficult for one of ordinary skill in the art” at the time of Appellant’s invention. Leapfrog Enters., Inc. v. Fisher-Price, Inc., 485 F.3d 1157, 1162 (Fed. Cir. 2007) (citing KSR, 550 U.S. at 418). Thus, Appellant does not persuade us of error in the Examiner’s obviousness rejection of representative claim 1. Independent claims 17, 28, and 29 include limitations of commensurate scope (except as noted, supra). Appeal 2021-002555 Application 15/625,578 13 Therefore, we affirm the Examiner’s obviousness rejection of representative claim 1 and claims 17, 28, and 29 not separately argued with particularity (supra). See 37 C.F.R. § 41.37(c)(1)(iv). Claims 9 and 22 The Examiner rejects dependent claims 9 and 22 as being obvious in view of Bogojeska and Salakhutdinov. See Final Act. 15–16, 23–24; Ans. 12–13. Claim 9 recites, in pertinent part, “DNN parameters are learned via at least one bottom-up pass followed by at least one top-down pass; in the bottom-up pass, all hidden nodes of a hidden layer of the DNN are sampled in parallel; and in the top-down pass, all visible nodes are sampled in parallel.” Appeal Br. 24 (Claims App.). Appellant concedes that Salakhutdinov “mention[s] ‘bottom up and top down pass,’” but contends “Salakhutdinov does not mention” that “in the bottom-up pass, all hidden nodes of a hidden layer of the DNN are sampled in parallel; and in the top- down pass, all visible nodes are sampled in parallel.” Appeal Br. 15; see Reply Br. 11–12. The Examiner counters that: Salakhutdinov teaches these limitations, including training using contrastive divergence (sec. 2, p. 2), learning via at least one bottom-up pass followed by at least one top-down pass (sec. 3, p. 3), all hidden nodes on a bottom-up pass (sec. 3.1, p. 5) sampled (sec. 2, p. 2) in parallel (sec. 2.1, p. 3), all visible nodes in a top-down pass (sec. 3, p. 3) sampled (sec. 2, p. 2) in parallel (sec. 2.1, p. 3). Ans. 15. The Examiner’s broad citations to Salakhutdinov (supra) do not clearly explain how Salakhutdinov teaches that all the hidden nodes in the hidden layers are sampled in parallel. The Examiner has not sufficiently shown that Salakhutdinov teaches performing the parallel hidden node Appeal 2021-002555 Application 15/625,578 14 sampling as required by these claims. Consequently, we are constrained by the record before us to find that the Examiner erred in finding Bogojeska and Salakhutdinov render obvious Appellant’s claims 9 and 22. Thus, we do not sustain the Examiner’s obviousness rejection of claims 9 and 22. Claims 10, 11, and 23 The Examiner rejects dependent claims 10, 11, and 23 as being obvious in view of Bogojeska and Salakhutdinov. See Final Act. 16–17, 24; Ans. 13. Claim 10 recites, in pertinent part, “the DNN parameters are a first set of DNN parameters, and wherein the method further comprises receiving a second set of one or more DNN parameters to be applied to received inputs.” Appeal Br. 26 (Claims App.). Appellant contends that the claims require two distinct sets of parameters and Bogojeska does not teach two distinct sets of parameters being applied to the same inputs. See Appeal Br. 15–17; Reply Br. 12–13. Appellant’s arguments are persuasive of Examiner error. Claim 10 (as well as claims 11 and 23) requires two sets of parameters that can operate on the same inputs. As explained by Appellant, even if we accept the Examiner’s interpretation of Bogojeska, at best Bogojeska describes two distinct neural networks operating on the inputs, not two sets of parameters (in the same neural network). See Appeal Br. 15–17; Reply Br. 12–13. The Examiner has not sufficiently shown that Bogojeska teaches the two sets of parameters. Consequently, we are constrained by the record before us to find that the Examiner erred in finding Bogojeska and Salakhutdinov render obvious Appellant’s claims 10, 11, and 23, and we do not sustain the Examiner’s obviousness rejection of these claims. Appeal 2021-002555 Application 15/625,578 15 Claims 12 and 25 The Examiner rejects dependent claims 12 and 25 as being obvious in view of Bogojeska and Salakhutdinov. See Final Act. 17, 25; Ans. 13–15. Claim 12 recites, in pertinent part, “one or more of the DNNs in the first set of DNNs is implemented using circuitry in one or more of the processing units.” Appeal Br. 26 (Claims App.). Appellant contends, similar to claim 1 (supra), that Bogojeska doesn’t teach processing units (or processors), much less implementing DNNs in processing unit circuitry. See Appeal Br. 17– 18; Reply Br. 13. Appellant does not persuade us of error in the Examiner’s rejection of claims 12 and 25 for the same reasons as claim 1 (supra). Bogojeska discloses a computer system including a processor. See Bogojeska, Fig. 1 (elements 100, 105); Bogojeska ¶¶ 14–15. Accordingly, we affirm the Examiner’s rejection of claims 12 and 25. Claim 14 The Examiner rejects dependent claim 14 as being obvious in view of Bogojeska and Salakhutdinov. See Ans. 15–16. Claim 14 recites, in pertinent part, “the one or more DNNs are applied using a first core of a processing unit, and the at least one signal includes at least one control signal for a second core of the processing unit.” Appeal Br. 26 (Claims App.). Appellant contends that “Bogojeska certainly does not say anything about how different cores of a processing unit may be used to implement one or more DNNs.” Appeal Br. 18; see Reply Br. 14. Appellant’s arguments are persuasive of Examiner error. The Examiner does not address claim 14 in the Final Office Action and simply states in the Examiner’s Answer that “[d]ue to similarity of language of Appeal 2021-002555 Application 15/625,578 16 claims 13 and 14, the limitations of claim 14 are rejected under the same reasons as claim 13.” Ans. 15. Although claims 13 and 14 recite similar features, claim 14 requires a multi-core processor and using one (first) core to implement the DNNs. The Examiner has not sufficiently shown that Bogojeska teaches using one core of a multi-core processor to implement the DNNs. Consequently, we are constrained by the record before us to find that the Examiner erred in finding Bogojeska and Salakhutdinov render obvious Appellant’s claim 14, and we do not sustain the Examiner’s obviousness rejection of claim 14. Claim 21 The Examiner rejects dependent claim 21 as being obvious in view of Bogojeska and Salakhutdinov. See Final Act. 23; Ans. 16. Claim 21 recites, in pertinent part, “the system is further configured to provide a warning signal . . . if one or more unexpected branches, faults, or exceptions in one or both of the first and second processing units is predicted.” Appeal Br. 28 (Claims App.). Appellant contends the Examiner’s determinations that the predictive modeler may predict whether a feature is problematic or unproblematic “does not arrive at the specific limitation of claim 21,” i.e., “a warning . . . if one or more unexpected branches, faults, or exceptions in one or both of the first and second processing units is predicted,” and that “[t]here is no discussion at all in Bogojeska of ‘unexpected branches, faults, or exceptions.” Appeal Br. 19; see Reply Br. 15. We agree with Appellant’s arguments. The Examiner does not explain in any detail how Bogojeska describes the delineated predictions that trigger the warning. The Examiner has not sufficiently shown that Appeal 2021-002555 Application 15/625,578 17 Bogojeska teaches the disputed predictions or the DDN output including a warning. Consequently, we are constrained by the record before us to find that the Examiner erred in finding Bogojeska and Salakhutdinov render obvious Appellant’s claim 21. Appellant persuades us that the Examiner erred in rejecting claim 21, and we do not sustain the Examiner’s obviousness rejection of claim 21. Claims 2, 4, 13, 15, 18, 24, and 27 The Examiner rejects dependent claims 2, 4, 13, 15, 18, 24, and 27 as being obvious in view of Bogojeska and Salakhutdinov. See Final Act. 14– 15, 18–19, 22–23, and 25–26. Appellant does not separately argue the rejection of claims 2, 4, 13, 15, 18, 24, and 27 with particularity. See Appeal Br. 5–19. For the same reasons as claim 1, Appellant does not persuade us of error in the Examiner’s obviousness rejection of claims 2, 4, 13, 15, 18, 24, and 27, not separately argued with particularity. See 37 C.F.R. § 41.37(c)(1)(iv). Therefore, we affirm the Examiner’s obviousness rejection of claims 2, 4, 13, 15, 18, 24, and 27. Claims 3 and 5 The Examiner rejects dependent claims 3 and 5 as being obvious in view of Bogojeska, Salakhutdinov, and Hughes. See Final Act. 33–35; Ans. 17–20. Claim 3 recites, in pertinent part, “one or more of the DNNs in the first set of DNNs is implemented within a tool for creating an optimal processor design.” Appeal Br. 25 (Claims App.). Appellant contends, similar to claim 1 (supra), that the Examiner has not provided a sufficient rationale for combining Bogojeska, Salakhutdinov, and Hughes. See Appeal Br. 19–20; Reply Br. 15–16. Appeal 2021-002555 Application 15/625,578 18 Appellant does not persuade us of error in the Examiner’s rejection of claims 3 and 5 for the same reasons as claim 1 (supra). Appellant has not presented any persuasive evidence or argument that the Examiner’s rationale for combining Bogojeska, Salakhutdinov, and Hughes was inadequate. Accordingly, we affirm the Examiner’s rejection of claims 3 and 5. Claims 6–8, 19, and 20 The Examiner rejects dependent claims 6–8, 19, and 20 as being obvious in view of Bogojeska, Salakhutdinov, and Chedid. See Final Act. 35–39; Ans. 20–23. Appellant presents generalized arguments that the Examiner mis-construed or mis-quoted the claims and that Bogojeska, Salakhutdinov, and Chedid do not teach the various features of the claims. See Appeal Br. 20–22; Reply Br. 16–17. Appellant does not persuade us of error in the Examiner’s rejection of claims 6–8, 19, and 20 because Appellant has not presented any persuasive evidence or argument that the Examiner misconstrued the claims or that Chedid in combination with Bogojeska and Salakhutdinov would not have at least suggested the features of claims 6–8, 19, and 20. Accordingly, we affirm the Examiner’s rejection of claims 6–8, 19, and 20. Claims 16 and 26 The Examiner rejects dependent claims 16 and 26 as being obvious in view of Bogojeska, Salakhutdinov, and Sheikh. See Final Act. 39–41; Ans. 23–24. Appellant contends, with respect to claim 16, that “in Sheikh, the power level, not the temperature, is the input because a change in the power level results in a change in GPU temperature” and, therefore, “Sheikh does not suggest that a temperature of a processing unit can be received as an input to a DNN.” Appeal Br. 23; see Appeal Br. 22–23; Reply Br. 17. With Appeal 2021-002555 Application 15/625,578 19 respect to claim 26, Appellant also generally contends that Bogojeska, Salakhutdinov, and Sheikh are not properly combined or do not teach claim 26—“a general statement that Sheikh mentions a GPU is not sufficient . . . to combine Sheikh with Salakhutdinov and Bogojeska to arrive at claim 26.” Appeal Br. 23; see Reply Br. 17. Appellant does not persuade us of error in the Examiner’s rejection of claims 16 and 26 for the same reasons set forth by the Examiner. See Final Act. 39–41; Ans. 23–24. Appellant’s generalized attorney argument does not persuade us the Examiner erred in interpreting Sheikh or combining Sheikh with Bogojeska and Salakhutdinov. Therefore, we affirm the Examiner’s rejection of claims 16 and 26. CONCLUSION Appellant has not shown persuasively that the Examiner erred in rejecting claims 1–8, 12, 13, 15–20, and 24–29 under 35 U.S.C. § 103. Appellant has shown persuasively that the Examiner erred in rejecting claims 9–11, 14, and 21–23 under 35 U.S.C. § 103. We, therefore, sustain the Examiner’s rejections of claims 1–8, 12, 13, 15–20, and 24–29, but do not sustain the Examiner’s rejection of claims 9–11, 14, and 21–23. Appeal 2021-002555 Application 15/625,578 20 DECISION SUMMARY In summary: Claim(s) Rejected 35 U.S.C. § Reference(s)/Basis Affirmed Reversed 1, 2, 4, 9– 15, 17, 18, 21–25, 27– 29 103 Bogojeska, Salakhutdinov 1, 2, 4, 12, 13, 15, 17, 18, 24, 25, 27–29 9–11, 14, 21–23 3, 5 103 Bogojeska, Salakhutdinov, Hughes 3, 5 6–8, 19, 20 103 Bogojeska, Salakhutdinov, Chedid 6–8, 19, 20 16, 26 103 Bogojeska, Salakhutdinov, Sheikh 16, 26 Overall Outcome 1–8, 12, 13, 15–20, 24– 29 9–11, 14, 21–23 RESPONSE No time period for taking any subsequent action in connection with this appeal may be extended under 37 C.F.R. § 1.136(a)(1)(iv). See 37 C.F.R. § 41.50(f). AFFIRMED-IN-PART Copy with citationCopy as parenthetical citation