UNITED STATES PATENT AND TRADEMARK OFFICE
`
`____________
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`
`____________
`
`TOYOTA MOTOR CORPORATION,
`
`Petitioner
`
`
`
`v.
`
`
`
`AMERICAN VEHICULAR SCIENCES,
`
`Patent Owner
`
`
`
`Patent No. 5,845,000
`
`Issue Date: December 1, 1998
`
`Title: OPTICAL IDENTIFICATION AND MONITORING SYSTEM USING
`PATTERN RECOGNITION FOR USE WITH VEHICLES
`
`____________
`
`
`
`PETITIONER’S REPLY TO PATENT OWNER’S RESPONSE
`
`Case No. IPR2013-00424
`
`
`
`

`

`TABLE OF CONTENTS
`
`
`Page
`
`
`I.
`II.
`
`INTRODUCTION ...................................................................................................... 1
`THE “GENERATED FROM” LANGUAGE IS NOT A
`LIMITATION OF CHALLENGED CLAIMS 10, 11, 16, 17, 19, AND
`20 FOR PURPOSES OF THE PATENTABILITY ANALYSIS ....................... 3
`III. THE “GENERATED FROM” LANGUAGE OF CHALLENGED
`CLAIMS 10, 11, 16, 17, 19, 20, AND 23 IS NOT LIMITED TO
`TRAINING WITH REAL DATA ........................................................................... 4
`IV. LEMELSON EXPLICITLY DISCLOSES TRAINING WITH REAL
`DATA ............................................................................................................................. 6
`A.
`Lemelson Discloses Training With All Types of “Known Inputs,”
`Including “Real Data” ...................................................................................... 6
`B. One of Ordinary Skill Would Have Understood Lemelson’s
`Disclosure of “Known Inputs” to Refer to Training with Real
`Data ..................................................................................................................... 7
`1.
`Simulated Data ....................................................................................... 8
`2.
`Partial Data ............................................................................................. 9
`Lemelson Separately Discloses “Adaptive Operation” and “On-
`Line Adjustment” of its Neural Network Which Constitutes
`Training with “Real Data” .............................................................................10
`D. Dr. Koutsougeras’s Declaration Should Be Given Little Weight
`Because He Lacks Expertise With Neural Networks in Vehicles ...........10
`V. GROUND OF REVIEW 1: CLAIMS 10, 11, 19, AND 23 ARE
`ANTICIPATED UNDER 35 U.S.C. § 102(E) BY LEMELSON .....................11
`VI. GROUND OF REVIEW 2: CLAIMS 10, 11, 19 AND 23 ARE
`OBVIOUS UNDER 35 U.S.C. § 103(A) OVER LEMELSON AND
`ASAYAMA ..................................................................................................................11
`VII. GROUND OF REVIEW 3: CLAIMS 16, 17, AND 20 ARE OBVIOUS
`UNDER 35 U.S.C. § 103(A) OVER LEMELSON AND YANAGAWA .......12
`A.
`Lemelson Discloses the “Generated From” Language, Which
`Would Also Have Been Obvious to One of Ordinary Skill .....................12
`AVS is Wrong that One of Ordinary Skill Would Not have Tried
`to Improve Upon the System of Yanagawa Using a Neural
`Network ............................................................................................................13
`
`C.
`
`B.
`
`
`
`
`
`-i-
`
`
`
`

`

`TABLE OF CONTENTS
`(continued)
`
`Page
`
`C.
`
`The Neural Network of Lemelson has Speed and Reliability
`Advantages Over Traditional Computational Methods ............................14
`
`-ii-
`
`
`
`
`
`
`
`
`
`
`
`
`
`

`

`TABLE OF AUTHORITIES
`
`
`Page
`
`
`Cases
`Ex Parte Klasing et al.,
`App. No. 11/507,120, 2013 Pat. App. LEXIS 1619 (PTAB March 14, 2013 ............. 4
`Greenliant Sys.,
`Inc. v. Xicor LLC, 692 F.3d 1261 (Fed. Cir. 2012) ............................................................. 4
`In re Baxter Travelnol Labs.,
`952 F.2d 388 (Fed Cir. 1991) ............................................................................................... 8
`In re Gleave,
`560 F.3d 1331 (Fed Cir. 2009) ............................................................................................. 7
`In re Graves,
`69 F.3d 1147 (Fed. Cir. 1996) .............................................................................................. 8
`In re Petering,
`301 F.2d 676 (C.C.P.A. 1962) .............................................................................................. 7
`In re Warmerdam,
`33 F.3d 1354 (Fed. Cir. 1994) ......................................................................................... 2, 4
`SmithKline Beecham Corp. v. Apotex Corp.,
`439 F.3d 1312 (Fed. Cir. 2006) ....................................................................................... 1, 4
` Statutes
`37 C.F.R. § 42.23 ....................................................................................................................... 1
`
`37 C.F.R. § 42.24 ....................................................................................................................... 1
`
`35 U.S.C. § 102 ......................................................................................................................... 11
`
`35 U.S.C. § 103 .................................................................................................................. 11, 12
`
`
`
`
`
`-iii-
`
`
`
`

`

`
`I.
`
`INTRODUCTION
`
`IPR2013-00424
`
`Petitioner Toyota Motor Corporation (“TMC”) submits this Reply under 37
`
`C.F.R. § 42.23-24 to Patent Owner’s Response (Paper 29) in IPR2013-00424
`
`concerning U.S. Patent No. 5,845,000 (“the ’000 patent”). This filing is timely. See
`
`Papers 17 (Scheduling Order) and 26 (Stipulation to Adjust Schedule).
`
`AVS argues that U.S. Patent No. 6,553,130 (Ex. 1002, “Lemelson”) does not
`
`disclose either a pattern recognition algorithm “generated from data of possible
`
`exterior objects and patterns of received electromagnetic illumination from the
`
`possible exterior objects,” as required by claims 10 and 23, or the materially indistinct
`
`language in claim 16 (hereinafter referred to individually or collectively as the
`
`“generated from” language). AVS asserts that this language requires training with data
`
`and waves from actual objects (hereinafter, “real data”), as opposed to simulated data
`
`and waves (hereinafter, “simulated data”) or “data and waves not representing exterior
`
`objects to be detected” (hereinafter, “partial data”). AVS also asserts that Lemelson’s
`
`disclosure of training is too vague to discern which of the three categories of data
`
`(real, simulated, or partial) is taught. AVS asserts that Petitioner and the Board must,
`
`therefore, have implicitly been relying on the doctrine of inherency. AVS is wrong.
`
`First, the “generated from” language is not a limitation in claims 10 or 16,
`
`because it is a process step within apparatus claims. See SmithKline Beecham Corp. v.
`
`Apotex Corp., 439 F.3d 1312, 1317, 1319 (Fed. Cir. 2006) (“one cannot avoid
`
`anticipation by an earlier product disclosure by claiming the same product more
`
`-1-
`
`

`

`
`narrowly, that is, by claiming the product as produced by a particular process.”); In re
`
`IPR2013-00424
`
`Warmerdam, 33 F.3d 1354, 1360-61 & n. 6 (Fed. Cir. 1994) (noting that the language
`
`“data representing a bubble hierarchy generated by the method of . . .” likely fit into
`
`the “conventional definition” of a product-by-process claim). The claimed “pattern
`
`recognition algorithm” constitutes computer code, regardless of how it was created.
`
`AVS does not argue that generating it with real data structurally alters the code.
`
`Second, even if the “generated from” language constitutes a limitation, it is not
`
`limited to training with real data. The claims merely specify that the algorithm is
`
`generated from (1) data of possible exterior objects, and (2) patterns of received
`
`electromagnetic illumination from those possible exterior objects. The claimed
`
`patterns “of” received electromagnetic illumination, as opposed to, “patterns from”
`
`electromagnetic illumination, merely require patterns representing what received
`
`electromagnetic illumination would look like (which would include simulations).
`
`Third, Lemelson explicitly discloses the “generated from” language, even under
`
`AVS’s construction. Lemelson discloses a neural network trained to identify roadway
`
`hazards, such as automobiles and pedestrians, by providing “known inputs” until
`
`desired output responses are obtained. Ex. 1002 at 8:1-10. AVS’s expert, Dr.
`
`Koutsougeras, admits that real data was one “known input” at the time of Lemelson,
`
`and that there were only two other categories of data that could have been used
`
`(simulated or partial data). Ex. 1019 at 86:25-87:14, 163:18-164:7. The disclosure of
`
`“known inputs” would, therefore, have been sufficient to connote to one of ordinary
`
`-2-
`
`

`

`
`skill that any known category of data (i.e., real, simulated, or partial) could be used for
`
`IPR2013-00424
`
`training. In any event, as explained by TMC’s expert, Dr. Papanikolopoulos, one of
`
`ordinary skill would have read Lemelson to refer to real data because neither
`
`simulated data nor partial data could have reasonably been used to identify exterior
`
`objects. Ex. 1020, Reply Declaration of Dr. Papanikolopoulos at ¶¶ 10-12.
`
`Specifically with respect to Ground of Review 3, AVS further argues that it
`
`would not have been obvious to use the neural network of Lemelson with Yanagawa
`
`because Yanagawa’s headlight dimming system was already functional with simple
`
`mathematical equations. However, Yanagawa only discloses a method for identifying
`
`dual-light preceding vehicles that emit either white or red light; it does not explain
`
`how to handle, for example, motorcycles, street lamps, or police cars with flashing red
`
`lights. There is little dispute that Lemelson’s neural network could have dealt with
`
`these additional problems, and would have also offered improved speed, reliability
`
`and computational robustness. One of ordinary skill would, therefore, have used the
`
`neural network of Lemelson to improve the headlight dimming system of Yanagawa.
`
`II. THE “GENERATED FROM” LANGUAGE IS NOT A LIMITATION
`OF CHALLENGED CLAIMS 10, 11, 16, 17, 19, AND 20 FOR
`PURPOSES OF THE PATENTABILITY ANALYSIS
`
`Independent claims 10 and 16 (and the challenged claims depending therefrom)
`
`are directed to a “system” and require a “pattern recognition algorithm” embodied as
`
`computer code on a processor. AVS asserts that (i) Lemelson could have generated
`
`the algorithm using any of three categories of data, each of which would have been
`
`-3-
`
`

`

`
`sufficient to train the neural network, and that (ii) this is insufficient to meet the
`
`IPR2013-00424
`
`“generated from” language. Paper 29 at 18-21. But, under this framework, the
`
`“generated from” language in claims 10, 11, 16, 17, 19, and 20 is not a limitation for
`
`purposes of patentability because it merely specifies the method of creating the
`
`algorithm. See SmithKline Beecham Corp., 439 F.3d at 1317. The “generated from”
`
`language does not serve to limit the structural elements in any way. See Greenliant Sys.,
`
`Inc. v. Xicor LLC, 692 F.3d 1261, 1268 (Fed. Cir. 2012); Warmerdam, 33 F.3d at 1360-
`
`61, fn. 6 (holding that a software claim including the language “generated by” was
`
`likely in product-by-process format); see also Ex Parte Klasing et al., App. No.
`
`11/507,120, 2013 Pat. App. LEXIS 1619, at *8-10 (PTAB March 14, 2013) (holding
`
`that the term “one-piece metal casting” was in product-by-process format and did not
`
`limit the claim for purposes of patentability.) In fact, AVS makes no argument that a
`
`pattern recognition algorithm is different depending upon how it was generated, and,
`
`actually, argues the opposite. Paper 29 at 18-21. The “generated from” language is,
`
`therefore, not a limitation for purposes of patentability.
`
`III. THE “GENERATED FROM” LANGUAGE OF CHALLENGED
`CLAIMS 10, 11, 16, 17, 19, 20, AND 23 IS NOT LIMITED TO
`TRAINING WITH REAL DATA
`
`Even if it is a limitation, the “generated from” language does not require
`
`training with real data, as AVS argues. See Paper 29 at 12-14; Ex. 2002 at ¶¶ 43-45,
`
`54-55. Claim 10 is representative and requires an “algorithm generated from data of
`
`possible exterior objects and patterns of received electromagnetic illumination from
`
`-4-
`
`

`

`
`the possible exterior objects.” The parties’ dispute revolves around the term “of” and
`
`IPR2013-00424
`
`whether, as AVS would have it, “data of” and “patterns of” require the claimed
`
`algorithm to be generated from real data, as opposed to simulated or partial data.
`
`Petitioner submits that the term “of” is not so restrictive.
`
`The “of” in “data of possible exterior objects” and “patterns of received
`
`electromagnetic illumination” merely requires that the “data” and “patterns” used to
`
`generate the claimed algorithm represent “possible exterior objects,” and “received
`
`electromagnetic illumination,” respectively. This language does not exclude simulated
`
`data or patterns. By analogy, an image “of” a car might be a digital image file, a
`
`printed analog photo, a CAD drawing, or a painting. Notably, the claims do not use
`
`the more specific preposition “from,” as in “data from possible exterior objects”—
`
`even though AVS’ expert’s declaration (as well as AVS’s Response) is explicitly based
`
`on that misleading paraphrasing of the claim language:
`
`The phrase “known inputs” could mean numerous things other
`
`than data from possible exterior objects and patterns of received
`
`electromagnetic illumination from the possible exterior objects.
`
`Ex. 2002 at ¶ 56 (emphasis added); see id. at ¶ 57, 58, 64; Paper 29 at 17.
`
`
`
`AVS is also wrong that training with partial data (such as data of license plates)
`
`falls outside the scope of the “generated from” language. A license plate is an
`
`“exterior object(s)” in its own right. See, e.g., IPR2013-00419, Paper 19 at 12-14; Ex.
`
`1019 at 166:14-170:14. And, there is no claim requirement that training be performed
`
`-5-
`
`

`

`
`on the same exterior object that is later identified. Training is performed on
`
`IPR2013-00424
`
`“possible exterior objects,” while the identification is of an exterior object that
`
`presents a hazard to the vehicle.
`
`IV. LEMELSON EXPLICITLY DISCLOSES TRAINING WITH REAL
`DATA
`
`Even if AVS were correct on claim construction, Lemelson discloses training
`
`with real data and therefore would still meet the “generated from” language.
`
`A.
`
`Lemelson Discloses Training With All Types of “Known Inputs,”
`Including “Real Data”
`
`AVS asserts that there are three categories of data that could have been used to
`
`train the Lemelson neural network (real, simulated and partial), and that only training
`
`with real data meets the “generated from” language. See Paper 29 at 14-21; Ex. 2002
`
`at ¶¶ 56-70; Ex. 1019 at 89:15-25. But, even if AVS is correct, Lemelson still discloses
`
`the “generated from” language. AVS is incorrect in its assertion that, because
`
`Lemelson teaches training with “known inputs,” as opposed to explicitly using the
`
`words “real data” or “real objects,” it does not explicitly meet the claim limitation.
`
` AVS’s argument essentially is that the question before the Board is an
`
`inherency dispute, i.e., that because “known inputs” could allegedly mean several
`
`things, the Board must consider whether training would have “necessarily” been
`
`performed with “real data.” See Paper 29 at 14-16. But, this is not a situation in
`
`which a reference is wholly silent as to a particular limitation and the doctrine of
`
`inherency is raised as a mechanism to fill the gap. Rather, as Dr. Papanikolopoulos
`
`-6-
`
`

`

`
`explains, one of ordinary skill in the art would have understood that “known inputs”
`
`IPR2013-00424
`
`include real data. Ex. 1020 at ¶¶ 10-12. Even AVS’s expert admits that: (i) “known
`
`inputs” would have had some meaning to one of ordinary skill; (ii) it would have
`
`referred to some category of data; (iii) training with real data was one of those known
`
`categories at the time of Lemelson; and (iv) it would not have been unreasonable to
`
`think that a skilled artisan reading Lemelson’s disclosure would have thought that
`
`“known inputs” referred to training with real data. See Ex. 1019 at 132:24-138:5,
`
`157:12-159:14, 163:18-164:7.
`
`There is no requirement, for anticipation, that Lemelson separately list the
`
`alleged three categories of training data; the shorthand phrase “known inputs” was
`
`sufficient. This situation is analogous to the genus/species context, where “known
`
`inputs” is the genus and “real data” is the claimed species (assuming AVS’s claim
`
`construction is correct). Even if AVS is correct that there potentially are three such
`
`categories of data (i.e., three possible species) within the “known input” genus, and
`
`that any of the categories could have been used for training, the “generated from”
`
`language is still met by Lemelson because the size of the genus is small and one of
`
`ordinary skill can at once envisage the claimed species. See In re Gleave, 560 F.3d 1331,
`
`1337-38 (Fed Cir. 2009); In re Petering, 301 F.2d 676, 681-82 (C.C.P.A. 1962).
`
`B. One of Ordinary Skill Would Have Understood Lemelson’s
`Disclosure of “Known Inputs” to Refer to Training with Real Data
`
`AVS asserts that Lemelson’s “known inputs” disclosure could have been
`
`-7-
`
`

`

`
`referring to training using simulated or partial data, instead of real data. But, as Dr.
`
`IPR2013-00424
`
`Papanikolopoulos explains, one of ordinary skill would have understood that training
`
`with simulated or partial data would have had little use in the context of training a
`
`neural network to identify exterior objects or radiation sources. AVS’s argument is
`
`based entirely on the declaration of Dr. Koutsougeras, who has no experience with
`
`neural networks in vehicles. See Section IV.D. below.
`
`Lemelson discloses a collision avoidance system, wherein a neural network is
`
`used to identify different types of potential roadway hazards, including, for example,
`
`road barriers, trucks, automobiles, pedestrians, signs and symbols. Ex. 1002 at 5:41-
`
`43; 8:1-6. The Lemelson neural network is specifically trained “with data identifying
`
`the potential roadway hazards,” (IPR2013-00419, Paper 19 at 20), which include, for
`
`example, “automobiles, trucks and pedestrians.” Ex. 1002 at 8:1-8. It then can
`
`control vehicle systems such as headlights, in response thereto. E.g., Ex. 1002 at 5:56-
`
`59. One of ordinary skill would have understood that training a neural network to
`
`identify exterior objects or sources of radiation (in the context of Lemelson) would
`
`have been done with real data, and not with simulated or partial data. Ex. 1020 at ¶¶
`
`10-24; In re Baxter Travelnol Labs., 952 F.2d 388, 390 (Fed Cir. 1991); In re Graves, 69
`
`F.3d 1147, 1152-53 (Fed. Cir. 1996).
`
`1.
`
`Simulated Data
`
`AVS and Dr. Koutsougeras rely upon a single reference—U.S. Patent No.
`
`5,537,327—in support of the argument that Lemelson may have been referring to
`
`-8-
`
`

`

`
`training with simulated data. But the ’327 patent relates to using neural networks to
`
`IPR2013-00424
`
`identify fault impedances in electrical power systems; it involves none of the same
`
`complications involved in identifying all possible exterior objects that could collide
`
`with a vehicle. Ex. 1020 at ¶ 22. Moreover, AVS and its expert fail to appreciate the
`
`difficulties that one of ordinary skill in the art would have encountered when
`
`attempting to train a system with simulated data. For example, training with simulated
`
`data would have required complicated equipment not contemplated by Lemelson,
`
`such as prohibitively complex training data generators. Ex. 1020 at ¶¶ 18-21.
`
`2.
`
`Partial Data
`
`One of ordinary skill would have understood that partial data would have been
`
`similarly useless for neural network training. Ex. 1020 at ¶¶ 13-17. Dr. Koutsougeras
`
`gives just one example of partial data that Lemelson could have been using: license
`
`plates. See Ex. 2001 at ¶ 70. But, training with license plate data could only be useful
`
`for identification of vehicles, and not for the variety of objects identified by the
`
`Lemelson neural network. Ex. 1002 at 5:41-43; 8:1-6. Even Dr. Koutsougeras admits
`
`that a neural network could not be trained to identify pedestrians (as performed by
`
`the system of Lemelson) using data of license plates. Ex. 1019 at 176:4-21. Training
`
`using partial data may be useful when there is only a single type of exterior object that
`
`is of interest, but not when there are hundreds of potentially different hazardous
`
`-9-
`
`

`

`
`objects, as in Lemelson.1 Ex. 1020 at ¶¶ 13-17.
`
`IPR2013-00424
`
`In sum, one of ordinary skill in the art would have understood that “known
`
`inputs” in Lemelson refers to real data, and not to simulated or partial data.
`
`C.
`
`Lemelson Separately Discloses “Adaptive Operation” and “On-
`Line Adjustment” of its Neural Network Which Constitutes
`Training with “Real Data”
`
`Lemelson also discloses real-world, post-laboratory training: “Adaptive
`
`operation is also possible with on-line adjustment of network weights to meet imaging
`
`requirements.” Ex. 1002 at 8:9-10; see also Figs. 1 and 2 (showing the TV camera as an
`
`input into the image analyzing computer, which stores the neural network).
`
`D. Dr. Koutsougeras’s Declaration Should Be Given Little Weight
`Because He Lacks Expertise With Neural Networks in Vehicles
`
`As discussed above, AVS relies on a declaration from Dr. Koutsougeras for the
`
`propositions that Lemelson’s disclosure is too vague to meet the “generated from”
`
`language, and that simulated or partial data could have been used for training.2
`
`
`1
`Nor would one of ordinary skill generally have believed that partial data would
`
`be useful to train a neural network. Ex. 1020 at ¶¶ 13-17.
`
`2
`
`Although AVS and Dr. Koutsougeras argue that Lemelson’s disclosure of
`
`training is vague, the inventor of the ’000 patent used a nearly verbatim disclosure in
`
`another one of his patents: “Neural networks used in the accident avoidance system
`
`of this invention are trained to recognize roadway hazards including automobiles,
`
`-10-
`
`

`

`
`However, his testimony should be given little weight because his neural network
`
`IPR2013-00424
`
`experience is almost exclusively related to handwriting recognition. See Ex. 2002 at
`
`48-49 (unpaginated C.V.). He has never published a paper on pattern recognition in
`
`vehicles, see Ex. 1019 at 39:10-22, and did not do any work with pattern recognition
`
`algorithms in vehicles until at least 2004 see id. at 26:14-20. Dr. Koutsougeras’s only
`
`purported experience with vehicle exterior monitoring systems was a 2004
`
`autonomous vehicle competition with a team of students; but, this project—unlike
`
`Lemelson and the ’000 patent—did not employ pattern recognition or a neural
`
`network, or even identify exterior objects. Id. at 30:6-31:22.
`
`V. GROUND OF REVIEW 1: CLAIMS 10, 11, 19, AND 23 ARE
`ANTICIPATED UNDER 35 U.S.C. § 102(E) BY LEMELSON
`
`AVS argues only that Lemelson fails to disclose the “generated from” language.
`
`For the reasons set forth above, this is wrong. The Board should therefore hold the
`
`claims in Ground of Review 1 to be unpatentable.
`
`VI. GROUND OF REVIEW 2: CLAIMS 10, 11, 19 AND 23 ARE OBVIOUS
`UNDER 35 U.S.C. § 103(A) OVER LEMELSON AND ASAYAMA
`
`AVS argues only that Lemelson fails to disclose the “generated from” language.
`
`For the reasons set forth above, this is wrong. The Board should therefore hold the
`
`claims in Ground of Review 2 to be unpatentable.
`
`
`trucks, animals and pedestrians. Training involves providing known inputs to the
`
`network resulting in desired output responses.” Ex. 1019 at 50:30-38.
`
`-11-
`
`

`

`
`VII. GROUND OF REVIEW 3: CLAIMS 16, 17, AND 20 ARE OBVIOUS
`UNDER 35 U.S.C. § 103(A) OVER LEMELSON AND YANAGAWA
`
`IPR2013-00424
`
`AVS argues that the combination of Yanagawa and Lemelson does not render
`
`claims 16, 17, or 20 obvious because (1) neither reference discloses the “generated
`
`from” term and (2) one of ordinary skill in the art would not have been motivated to
`
`combine the neural network of Lemelson with the simple system of Yanagawa. AVS
`
`is wrong for the reasons set forth below.
`
`A.
`
`Lemelson Discloses the “Generated From” Language, Which
`Would Also Have Been Obvious to One of Ordinary Skill
`
`As set forth above, Lemelson explicitly discloses the “generated from”
`
`language of claim 16, even under AVS’s narrow construction.
`
`In any event, it would have been obvious to one of ordinary skill to generate a
`
`pattern recognition algorithm using real data for use in the Yanagawa system. The
`
`purpose of Yanagawa is to identify pairs of vehicle headlights or taillights in order to
`
`determine whether to dim the high beams. Toyota Petition (Paper 2) at 50-51; Ex.
`
`1009 at 1. As set forth in TMC’s petition, (pp. 58-59), as well as in Section VII.B.
`
`below, one of ordinary skill would have been motivated to use a neural network with
`
`the Yanagawa system because this would have allowed for identification of additional
`
`radiation sources, such as single-light motorcycle taillights. One of ordinary skill
`
`would have trained such a neural network with real data, because this would have
`
`been the most efficient and accurate way for the system to learn how to identify all
`
`possible radiation sources. See Ex. 1020 at ¶¶ 26-27. One of ordinary skill would not
`
`-12-
`
`

`

`
`have used simulated data, because this would have required complex generation of
`
`IPR2013-00424
`
`simulations of all possible radiation sources and the situations in which a vehicle
`
`would encounter them. Id. at ¶ 28. As for AVS’s arguments about “partial data,”
`
`these have no relevance to Ground 3. One of ordinary skill would not train a neural
`
`network to identify possible radiation sources using “license plate” data (see Paper 29
`
`at 20-21); and training using “taillight” data would clearly meet the claim limitation
`
`since it is a source of radiation in its own right.3
`
`B.
`
`AVS is Wrong that One of Ordinary Skill Would Not have Tried to
`Improve Upon the System of Yanagawa Using a Neural Network
`
`Yanagawa discloses a relatively simple headlight dimming system that is able to
`
`detect preceding taillights and headlights and dim the vehicle’s high beams as
`
`necessary. Ex. 1009 at 2-4. The system operates with a color filter and by detecting
`
`pairs of lights ahead, which are presumed to be transmitted by a vehicle. Id. at 3-4.
`
`Applying equations that factor in the estimated distance to the vehicle, the system is
`
`able to output a binary ON or OFF outcome. Id. at 4. AVS’s expert admitted that
`
`Yanagawa does not disclose any method for dealing with, for example, single light
`
`objects, such as motorcycles or vehicles with one burned out taillight. Ex. 1019 at
`
`267:17-263:14. AVS’s expert also stated that Yanagawa does not provide equations
`
`3
`Tellingly, neither AVS nor Dr. Koutsougeras points to “taillights” as an
`
`example of “partial data,” like they did in IPR2013-00419. See IPR2013-00419, Paper
`
`33 at 19-20 and IPR2013-00419, Ex. 2001 at ¶ 64.
`
`-13-
`
`

`

`
`for dealing with false-positive lighting from, for example, street lamps, traffic lights or
`
`IPR2013-00424
`
`billboard reflections, or for dealing with police cars, which might have both white
`
`headlights and flashing red lights. Ex. 1019 at 263:15-267:10.
`
`One of ordinary skill would have sought to expand the applicability of the
`
`Yanagawa system to allow it to accurately dim headlights in response to all types of
`
`potential radiation sources. Yanagawa is a 1987 reference (filed in 1985). Ex. 1009.
`
`Subsequently, one of ordinary skill would have recognized that a neural network
`
`would have allowed such improvement. As Dr. Papanikolopolous explained, “[a]s
`
`computers improved from the late 1980’s to the early 1990’s, neural networks were
`
`viewed as a viable alternative.” Ex. 1013 at ¶ 36; id. at ¶ 32 (listing 13 neural network
`
`papers from between 1991 and 1996). “Neural network methodologies, such as back-
`
`propagation, provided ways to quickly adapt to the rapidly evolving scenes that
`
`vehicles would encounter.” Id. at ¶ 37 (emphasis added).
`
`One of ordinary skill in the art would have recognized that a neural network
`
`(such as that in Lemelson) would have been an ideal choice for the headlight dimming
`
`system, particularly because (as Dr. Papanikolopolous explained) they are very
`
`effective when the choices in decision-making are simple, such as the binary decision
`
`for turning the headlights “on” or “off.” Ex. 1013 at ¶ 120.
`
`C. The Neural Network of Lemelson has Speed and Reliability
`Advantages Over Traditional Computational Methods
`
`AVS’s arguments about Ground of Review 3 are based on unsound statements
`
`-14-
`
`

`

`
`regarding the reliability and benefits of neural networks. For example, AVS’ expert
`
`IPR2013-00424
`
`ignores that a neural network system would continue to function in the case of a
`
`component failure (in some limited capacity), whereas a standard computer would
`
`cease to function altogether. Ex. 1013 at ¶ 121. He also erroneously characterizes the
`
`system of Lemelson as a serial rather than a parallel system. Figures 3 and 5 of
`
`Lemelson clearly show multiple processing elements “PE” as 63 in Figure 3 and 73 in
`
`Figure 5. Ex. 1002 at 7:47-56, 8:21-29. Nowhere does Lemelson limit its system to
`
`single CPUs or to a single general purpose computer. See Ex. 1002 at 6:65-8:29.
`
`VII. CONCLUSION
`
`For the reasons in TMC’s Petition for Inter Partes Review of U.S. Patent No.
`
`5,845,000 (Paper 2), for the reasons in the Board’s decision to institute an inter partes
`
`review (Paper 16), and for the additional reasons set forth above, the Board should
`
`maintain its opinion of unpatentability of claims 10, 11, 16, 17, 19, 20, and 23.
`
`
`
`/Matt Berkowitz/
`Matt Berkowitz (Reg. No. 57,215)
`A. Antony Pfeffer (Reg. No. 43,857)
`Thomas R. Makin
`Kenyon & Kenyon LLP
`One Broadway
`New York, NY 10004
`Tel: 212-425-7200
`Attorney for Petitioner,
`Toyota Motor Corporation
`
`-15-
`
`

`

`
`
`
`
`CERTIFICATE OF SERVICE
`
`The undersigned hereby confirms that the foregoing Petitioner’s Reply to
`
`Patent Owner’s Response and Exhibits 1019-1023 were served on June 2, 2014 via e-
`
`mail upon the following counsel of record for Patent Owner:
`
`Thomas J. Wimbiscus
`Scott P. McBride
`Christopher M. Scharff
`twimbiscus@mcandrews-ip.com
`smcbride@mcandrews-ip.com
`cscharff@mcandrews-ip.com
`AVS-IPR@mcandrews-ip.com
`MCANDREWS HELD & MALLOY, LTD.
`500 W. Madison St., 34th Floor
`Chicago, IL 60661
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`/Matt Berkowitz/___
`Matt Berkowitz (Reg. No. 57,215)
`Kenyon & Kenyon LLP
`One Broadway
`New York, NY 10004
`Tel: 212-425-7200
`Fax: 212-425-5288
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`