`
`______________________________
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`
`
`
`TOYOTA MOTOR CORPORATION
`
`Petitioner
`
`
`
`v.
`
`
`
`Patent of AMERICAN VEHICULAR SCIENCES
`
`Patent Owner
`
`
`
`Patent No. 6,772,057
`
`Issue Date: August 3, 2004
`
`Title: VEHICLE MONITORING SYSTEMS USING IMAGE PROCESSING
`
`
`
`PATENT OWNER’S RESPONSE
`PURSUANT TO 37 CFR § 42.120
`
`Case No. IPR2013-00419
`
`
`
`
`
`
`
`
`
`
`
`TABLE OF CONTENTS
`
`INTRODUCTION ........................................................................................... 1
`
`SUMMARY OF THE ’057 PATENT, SCOPE AND CONTENT OF
`THE PRIOR ART, AND LEVEL OF ORDINARY SKILL........................... 4
`
`I.
`
`II.
`
`III. GROUNDS FOR WHICH REVIEW HAS BEEN INSTITUTED ................. 8
`
`IV. CLAIM CONSTRUCTION ............................................................................ 9
`
`V.
`
`THE BOARD SHOULD CONFIRM VALIDITY OF CLAIM 1-4, 7-
`10, 31, 41, 56, 59-62, AND 64 OVER THE GROUNDS ASSERTED
`IN THE PETITION ....................................................................................... 10
`
`A. None of the References Raised In The Review Disclose a
`“Pattern Recognition Algorithm Generated From Data of
`Possible Exterior Objects and Patterns of Received Waves from
`the Possible Exterior Objects” (claims 1-4, 7-10, 31, 41, 56, 59-
`61, 62, 64) ............................................................................................ 10
`
`(1) Lemelson ................................................................................... 11
`
`a.
`
`b.
`
`c.
`
`d.
`
`Lemelson does not expressly disclose the claim
`limitation ......................................................................... 12
`
`The Board’s decision to grant review based on
`Lemelson relied on the doctrine of inherency ................ 12
`
`Lemelson does not inherently disclose the claim
`limitation—it could have involved generating the
`algorithm with simulated data ........................................ 14
`
`Lemelson does not inherently disclose the claim
`limitation—it also could have involved generating
`an algorithm with data and waves not representing
`exterior objects to be detected ........................................ 19
`
`
`
`i
`
`
`
`e.
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`Toyota’s expert’s belated attempt at his deposition
`to
`read extra disclosure
`into Lemelson
`is
`unavailing ....................................................................... 22
`
`(2) Borcherts ................................................................................... 25
`
`(3) Asayama .................................................................................... 26
`
`(4) Yamamura ................................................................................. 26
`
`(5) Other References Cited In the Petition But For Which
`Review Was Not Granted ......................................................... 27
`
`B. None of the Obviousness Grounds Raised In The Review Fix
`The Failure To Disclose a “Pattern Recognition Algorithm
`Generated From Data of Possible Exterior Objects and
`Patterns of Received Waves from the Possible Exterior
`Objects” (claims 4, 31, 59) .................................................................. 27
`
`VI. THE BOARD SHOULD CONFIRM VALIDITY OF CLAIM 30, 32-
`34, 37-39, AND 62 OVER THE GROUNDS ASSERTED IN THE
`PETITION ..................................................................................................... 29
`
`VII. CONCLUSION .............................................................................................. 32
`
`
`
`
`
`
`
`
`
`ii
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`
`TABLE OF AUTHORITIES
`
`Cases
`
`Microsoft Corp. v. Proxyconn, Inc.,
`Case IPR2012-00026 (PTAB, Feb. 19, 2014) ........................... 13, 28, 29, 32
`
`Scaltech, Inc. v. Retec/Tetra, LLC.,
`178 F.3d 1378 (Fed. Cir. 1999) ..................................................................... 13
`
`Transclean Corp. v. Bridgewood Servs., Inc.,
`290 F.3d 1364 (Fed. Cir. 2002) ..................................................................... 13
`
`Verdegaal Bros. v. Union Oil Co. of California,
`814 F.2d 628 (Fed. Cir. 1987) ....................................................................... 11
`
`Statutes
`35 U.S.C. §314 ............................................................................................... 27
`Rules
`37 CFR §42.120 ..................................................................................... 1, 9, 27
`
`
`
`
`iii
`
`
`
`I.
`
`INTRODUCTION
`
`Patent Owner American Vehicular Sciences (“AVS”) submits the following
`
`response under 37 CFR §42.120 to the Petition filed by Toyota Motor Corporation
`
`(“Toyota”) requesting inter partes review of certain claims of U.S. Pat. No.
`
`6,772,057 (“the ‘057 patent”). This filing is timely pursuant to the Board’s
`
`Scheduling Order and the parties’ stipulation extending the deadline to March 20,
`
`2014. (See Paper 20, Scheduling Order (“The parties may stipulate to different
`
`dates for DUE DATES 1 through 3 (earlier or later, but no later than DUE DATE
`
`4).”); Paper 30, Notice of Stipulation).)
`
`
`
`AVS respectfully submits that the arguments presented and the additional
`
`evidence submitted, such as testimony from AVS expert Professor Cris
`
`Koutsougeras, PhD, show that at least claims 1-4, 7-10, 31, 41, 56, 59-62, and 64
`
`of the ‘057 patent are not anticipated or obvious in view of the grounds for review.
`
`AVS also reiterates the arguments with respect to claims 30, 32-34 and 37-39.
`
`Specifically, none of the prior art raised in the grounds for review discloses a
`
`key requirement in claims 1-4, 7-10, 31, 41, 56, 59-62, and 64 of the ‘057 patent—
`
`a “pattern recognition algorithm generated from data of possible exterior objects
`
`and patterns of received waves from the possible exterior objects.” (See Exhibit
`
`1001, ‘057 patent at independent claims 1 and 56 and dependent claims 31 and 41
`
`(emphasis added).) In other words, these claims require a pattern recognition
`
`
`
`1
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`algorithm that must be generated in this particular way. Toyota and its expert
`
`glossed over this claim requirement, suggesting that just any pattern recognition
`
`algorithm would suffice. But as AVS’s expert explains, and illustrates with
`
`evidentiary support, there are numerous different ways that a pattern recognition
`
`algorithm can be generated that would not satisfy this claim limitation.
`
`Importantly, Toyota and its expert had only alleged that two out of the eight
`
`prior art references that it asserted in its Petition (Lemelson and Pomerleau) even
`
`disclosed a “pattern recognition algorithm” at all (much less one generated as
`
`required by the above-listed ‘057 patent claims). (See Paper 1, Toyota’s Petition at
`
`10-21, 40-46.)
`
`Out of those two references, the Board found that Pomerleau was not
`
`appropriate for an anticipation or obviousness ground for review, because it
`
`disclosed using a trained pattern recognition algorithm for detecting road lines—
`
`not “objects” as required by the claims. (See Paper 19, Board’s Decision to
`
`Institute Inter Parte Review (“Board Decision”) at pp. 34-37.) The Board therefore
`
`substantively denied review based on any ground premised on Pomerleau, and
`
`Pomerleau is therefore not at issue. (Id.)
`
`With respect to Lemelson, Toyota and its expert only pointed to a single
`
`sentence in Lemelson that refers to how the pattern recognition algorithm is
`
`generated—a sentence that states that the training of Lemelson’s network involved
`
`
`
`2
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`“providing known inputs to the network resulting in desired output responses.”
`
`(See Paper 1, Toyota’s Petition at 11, citing Lemelson at 8:4-6.) Toyota glossed
`
`over the failure in Lemelson to disclose whether those “known inputs” included the
`
`specific inputs required by claims 1-4, 7-10, 31, 41, 56, 59-62, and 64.
`
`As discussed below, Toyota’s arguments, and the Board’s comments in
`
`response, implicitly rest on the doctrine of inherency. In other words, because
`
`Lemelson does not expressly disclose generating a trained pattern recognition
`
`algorithm “from data of possible exterior objects and patterns of received waves
`
`from the possible exterior objects,” in order to find anticipation, Toyota was
`
`required to show that Lemelson “necessarily” included that type of algorithm
`
`generation (i.e., not that it was merely possible or probable that Lemelson used the
`
`claimed type of algorithm generation). Toyota, however, did not establish this
`
`requirement, and could not establish this requirement, because there are in fact
`
`several types of “known inputs” that Lemelson could have been referring to other
`
`than the inputs required by the subject ‘057 patent claims.
`
`For example, Lemelson could have used simulated data to generate a pattern
`
`recognition algorithm, which would not involve “data of possible exterior objects
`
`and patterns of received waves from the possible exterior objects.” Or it could
`
`have used data or wave patterns relating to something other than “the possible
`
`exterior objects” for which the system is trying to provide a “classification,
`
`
`
`3
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`identification, or location.” For example, instead of training the system with data
`
`and patterns of received waves from cars, it could have involved training with
`
`images of license plates or tail-lights, which would fail to satisfy the claim.
`
`As such, the instituted grounds for review do not establish anticipation or
`
`obviousness of at least claims 1-4, 7-10, 30-34, 37-39, 41, 56, 59-62, and 64 of the
`
`‘057 patent. If the Board agrees that Lemelson does not “necessarily” disclose the
`
`claimed manner of generating an algorithm, then the instituted ground for review
`
`of claims 1-4, 7-10, 31, 41, 56, 59-62, and 64 based on anticipation by Lemelson
`
`fails, as do the instituted grounds for review of obviousness of claims 5, 31, and 59
`
`(in view of Lemelson in combination with Borchert or Asayama). And if the
`
`Board agrees with AVS regarding claim 30’s “on the rearview mirror”
`
`requirement, then claims 30, 32-34 and 37-39 also overcome the grounds for
`
`review. AVS requests that the Board confirm claims 1-4, 7-10, 30-34, 37-39, 41,
`
`56, 59-62, and 64.1
`
`II.
`
`SUMMARY OF THE ’057 PATENT, SCOPE AND CONTENT OF
`THE PRIOR ART, AND LEVEL OF ORDINARY SKILL
`
`The ‘057 patent claims generally relate to technology for monitoring an
`
`environment exterior of a vehicle, where the vehicle determines if any object is in
`
`1 AVS is not challenging the Board’s review of the remaining claims for which
`
`review was instituted, namely claims 40, 43, 46, 48, and 49.
`
`
`
`4
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`the path of vehicle, classifies or identifies the object, and affects other systems in
`
`the vehicle in response to the classification or identification of the object. But
`
`what made the ‘057 patent groundbreaking and superior to prior vehicle collision
`
`avoidance systems was the specific way that it implemented the system.
`
`Independent claims 1 and 56 (and dependent claims 31 and 41) recite a key aspect
`
`of the invention—a processor coupled to at least one receiver (e.g., an infrared
`
`receiver, CCD array, or radar receiver), where the processor implements a trained
`
`pattern recognition system (such as a neural network) that is trained with data and
`
`patterns of received waves from possible exterior objects.
`
`As AVS’s expert explains in his declaration, a pattern recognition system
`
`such as a neural network is fundamentally different than just a computer program.
`
`(Exhibit 2001, Koutsougeras Decl. at ¶ 15.) A computer program can be used if a
`
`programmer can guarantee knowing all possible variables. (Id.) But in an object
`
`detection system, this can be very difficult. (Id. at ¶¶ 15-16.) If the goal is to have
`
`the system detect whether an object is a car, it would be difficult to program such a
`
`system to compare a received image of a car to a database of images of all possible
`
`car models, in all possible colors, from all possible angles. (Id. at ¶ 18.)
`
`For that reason, the inventor of the ‘057 patent developed a way to perform
`
`this object recognition using a “pattern recognition algorithm” such as a neural
`
`network. (Id. at ¶¶ 16-20.) A pattern recognition algorithm does not just compare
`
`
`
`5
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`detected car to a database to find a match. Rather, it calculates degrees of
`
`similarity between something it has been told (or “trained”) is a car, versus
`
`something it has been told is not a car. (Id. at ¶ 18.) The more positive and
`
`negative examples (the “training set”) that the system is given, the more accurate it
`
`will be. (Id.)
`
`The inventor of the ‘057 patent also found that a specific type of training to
`
`generate the “training set” was the most effective. (See id. at ¶¶ 19, 20, 53.) The
`
`inventor disclosed and claimed generating the algorithm from “data of possible
`
`exterior objects and patterns of received waves from the possible exterior objects.”
`
`(Id.) For example, if the vehicle uses a radar receiver, a neural network could be
`
`trained with examples of received radar waves from possible objects such as cars,
`
`motorcycles, trucks, etc. (“i.e., “patterns or received waves from the possible
`
`exterior objects”), plus labels indicating the classification and possibly other
`
`information relating to the example object (i.e., “data”). (Id. at ¶¶ 19-20.) The
`
`examples of received radar waves from possible objects used to generate the
`
`algorithm can, therefore, be real radar waves, so that the system knows how to
`
`recognize radar waves received from that same object or a similar one when the
`
`vehicle is later driving down the road. (Id. at ¶ 20.) This can be done, for
`
`example, by putting actual examples of a possible object in front of a vehicle radar
`
`system, letting the system hit the object with radar waves that are thereafter
`
`
`
`6
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`received back by the system, and then telling the system the identity and
`
`classification of the object.2 (Id. at ¶ 20.) This is in contrast to other ways to train
`
`a pattern recognition system, such as through completely simulated data (a
`
`computer simulation of radar waves). (Id. at ¶¶ 46, 55-63.)
`
`As Professor Koutsougeras explains, therefore, the scope and content of the
`
`prior art to the ‘057 patent would have been narrower than that offered by Toyota
`
`and its expert, Dr. Papanikolopolous. (Id. at ¶ 34.) Professor Koutsougeras
`
`explains that the scope and content of the prior art would not have included
`
`generically any “vehicle sensing systems,” as there are many vehicle sensing
`
`systems that have no relevance or application to external object detection or pattern
`
`recognition systems. (Id.) Rather, the scope and content of the prior art would
`
`have included sensors and pattern recognition algorithms for object classification,
`
`including those for automotive use. (Id.) AVS and Professor Koutsougeras,
`
`however, do not have any fundamental disagreement with the definition of the
`
`level of ordinary skill proposed by Toyota and Dr. Papanikolopolous, and therefore
`
`have applied that definition of the level of ordinary skill for purposes of this IPR.
`
`
`2 This is not to say, of course, that every individual vehicle must be trained in this
`
`way. Once a single system has been trained, those saved examples of waves and
`
`label data can be transferred to other systems.
`
`
`
`7
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`III. GROUNDS FOR WHICH REVIEW HAS BEEN INSTITUTED
`
`Toyota’s Petition included sixteen proposed grounds for invalidity, based on
`
`eight different prior art references. (See Paper 19, Board Decision at pp. 6-7.) Of
`
`those sixteen proposed grounds, the Board granted review based on five of those
`
`grounds. Specifically, the Board granted review on the following grounds:
`
` Claims 1-4, 7-10, 40, 41, 46, 48, 49, 56, 59-61, and 64 as anticipated under
`
`35 U.S.C. § 102(e) by Lemelson;
`
` Claims 30-34, 37-39, and 62 for obviousness under 35 U.S.C. § 103(a) over
`
`Lemelson and Borcherts;
`
` Claims 4, 43, and 59 for obviousness under 35 U.S.C. § 103(a) over
`
`Lemelson and Asayama;
`
` Claim 34 for obviousness under 35 U.S.C. § 103(a) over Lemelson,
`
`Borcherts, and Asayama; and
`
` Claims 30, 32, and 37-39 for obviousness under 35 U.S.C. § 103(a) over
`
`Yamamura and Borcherts.
`
`(Paper 19, Board Decision at pp. 38-39.)
`
`Accordingly, claims 1-3, 7-10, 41, 56, 60, 61, and 64 only stand reviewed
`
`for alleged anticipation by Lemelson. Review of those claims was not instituted
`
`based on any other prior art reference, nor on any other ground. Claims 30-34, 37-
`
`39 and 62 stand reviewed only for alleged obviousness in view of Lemelson and
`
`
`
`8
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`Borcherts. Claims 4 and 59 stand reviewed only for alleged anticipation by
`
`Lemelson or alleged obviousness in view of Lemelson and Asayama.
`
`Pursuant to 37 CFR §42.120, AVS is addressing only the grounds for which
`
`review was instituted, for select claims. (See 37 CFR §42.120 (“A patent owner
`
`may file a response to the petition addressing any ground for unpatentabililty not
`
`already denied.”).)
`
`IV. CLAIM CONSTRUCTION
`
`For purposes of this IPR only, AVS does not contest the Board’s claim
`
`constructions. Any disagreements that AVS might have with the Board’s claim
`
`constructions are not material to the arguments in this Response.
`
`In particular, the Board provided the following constructions for the
`
`following terms:
`
`
`
` “trained pattern recognition algorithm” is construed as “an algorithm
`
`that processes a signal that is generated by an object, or is modified by interacting
`
`with an object, in order to determine to which one of a set of classes the object
`
`belongs, the algorithm having been taught, through a variety of examples, various
`
`patterns of received signals generated or modified by objects”;
`
`
`
`“trained pattern recognition means” is construed as “a neural
`
`computer or neural network trained for pattern recognition, and equivalents
`
`thereof”;
`
`
`
`9
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`“identify” / “identification” is construed as “determine that the
`
`
`
`object belongs to a particular set or class” and “identification” as “determination
`
`that the object belongs to a particular set or class”;
`
`
`
`“exterior object” is construed as “a material or physical thing outside
`
`the vehicle, not a part of the roadway on which the vehicle travels”;
`
`
`
`“rear view mirror” is construed as “a mirror that faces to the rear,
`
`which necessarily excludes non-rear-facing mirrors”; and
`
`
`
`“transmitter” is construed as “encompassing devices that transmit any
`
`type of electromagnetic waves, including visible light.”
`
`V. THE BOARD SHOULD CONFIRM VALIDITY OF CLAIM 1-4, 7-10,
`31, 41, 56, 59-62, AND 64 OVER THE GROUNDS ASSERTED IN THE
`PETITION
`
`A. None of the References Raised In The Review Disclose a “Pattern
`Recognition Algorithm Generated From Data of Possible Exterior
`Objects and Patterns of Received Waves from the Possible Exterior
`Objects” (claims 1-4, 7-10, 31, 41, 56, 59-61, 62, 64)
`
`As discussed, independent claims 1 and 56 and dependent claims 31 and 41
`
`require a specific type of training of the pattern recognition algorithm. These
`
`claims require a pattern recognition algorithm “generated from data of possible
`
`exterior objects and patterns of received waves from the possible exterior objects.”
`
`(See Exhibit 1001, ‘057 patent at claims 1, 56, 31, 41.)
`
`None of the references at issue in the instituted grounds for review (i.e.,
`
`
`
`10
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`Lemelson, Asayama, Borcherts, or Yamamura) disclose this claim limitation,
`
`either expressly or inherently. See Verdegaal Bros. v. Union Oil Co. of California,
`
`814 F.2d 628, 631 (Fed. Cir. 1987) (“A claim is anticipated only if each and
`
`every element as set forth in the claim is found, either expressly or inherently
`
`described, in a single prior art reference.”).
`
`(1) Lemelson
`
`The only reference that the Board found may disclose a “pattern recognition
`
`algorithm generated from data of possible exterior objects and patterns of received
`
`waves from the possible exterior objects” is Lemelson. (See Paper 19, Board’s
`
`Decision at pp. 15-21.) Review of claims 1-4, 7-10, 41, 56, 59-61, and 64 was
`
`instituted for anticipation by Lemelson. Review of claims 31 and 62 (which
`
`additionally require a receiver arranged on a rear view mirror) was only instituted
`
`for obviousness by Lemelson in view of Borcherts. And review of claims 4 and 59
`
`(which require an infrared transmitter) was also instituted for obviousness by
`
`Lemelson in view of Asayama.
`
`Lemelson, however, does not expressly disclose the nature and manner of
`
`how its neural network algorithm is generated, and it does not inherently (i.e.,
`
`“necessarily”) disclose that its neural network was generated as claimed.
`
`
`
`
`
`
`
`11
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`Lemelson does not expressly disclose the claim limitation
`
`Lemelson discloses a system for identifying objects exterior to a vehicle.
`
`a.
`
`(See Exhibit 1002, Lemelson at Abstract.) And it does disclose using a type of
`
`pattern recognition algorithm (a neural network) for identifying objects. (See
`
`Lemelson at 5:35-45.) The only discussion in Lemelson, however, relating to
`
`generating the neural network, merely states that “[t]raining involves providing
`
`known inputs to the network resulting in desired output responses.” (See Exhibit
`
`2001, Koutsougeras Decl. at ¶ 43, citing Lemelson at 8:4-6.)
`
`This is the only sentence from Lemelson that Toyota cited in its Petition as
`
`relating to the nature of Lemelson’s pattern recognition algorithm generation or
`
`training. (See Paper 1, Toyota’s Petition at p. 11.) And it is the only sentence that
`
`Toyota’s expert, Dr. Papanikolopolous, cites in his declaration with respect to how
`
`the trained pattern recognition algorithm in Lemelson is generated. (See Exhibit
`
`1016, Papanikolopoulos Decl. at ¶¶ 47-64.)
`
` Nowhere else
`
`in Dr.
`
`Papanikolopolous’s declaration does he allege that Lemelson discloses how its
`
`pattern recognition algorithm was generated. (See id.)
`
`b.
`
`The Board’s decision to grant review based on Lemelson
`relied on the doctrine of inherency
`
`
`The Board also did not rely on any express disclosure in Lemelson with
`
`respect to the “algorithm generated from” requirement of the subject ‘057 patent
`
`claims. The Board only found that Lemelson discloses training a neural network
`12
`
`
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`with “known inputs.” (Paper 19, Board Decision at 20.) From that, the Board only
`
`stated that “[i]t follows that, during training of the neural network, providing
`
`known inputs as disclosed in Lemelson involves providing the neural network with
`
`data identifying the potential roadway hazards, i.e., data of possible exterior
`
`objects, corresponding to the inputs, i.e., patterns of received waves from possible
`
`exterior objects.” (Id.) (emphasis added).
`
`The use by the Board of the phrases “it follows that” and “Lemelson
`
`involves” means that, although it did not reference the doctrine by name, the Board
`
`is applying the doctrine of inherency. Inherency, however, requires that a claimed
`
`limitation be “necessarily” and “inevitably” present. See Transclean Corp. v.
`
`Bridgewood Servs., Inc., 290 F.3d 1364, 1373 (Fed. Cir. 2002) (“Inherent”
`
`anticipation is appropriate only when the prior art necessarily includes a claim
`
`limitation that is not expressly disclosed.). It is not enough that a claim limitation
`
`was possibly or probably present in a prior art reference. See Scaltech, Inc. v.
`
`Retec/Tetra, LLC., 178 F.3d 1378, 1384 (Fed. Cir. 1999) (invalidity based on
`
`inherency is not established by mere “probabilities or possibilities”). See also, e.g.,
`
`Microsoft Corp. v. Proxyconn, Inc., Case IPR2012-00026 (PTAB, Feb. 19, 2014)
`
`(“A finding of anticipation by inherency requires more than probabilities or
`
`possibilities. Based on the evidence discussed above, it is possible to infer that
`
`Perlman describes such permanent storage memory. However, Microsoft has not
`
`
`
`13
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`presented evidence that the computers or routers described by Perlman necessarily
`
`use permanent storage memory as recited in claims 1 and 3.”).
`
`Here, the only way that Lemelson could inherently disclose a “pattern
`
`recognition algorithm generated from data of possible exterior objects and patterns
`
`of received waves from the possible exterior objects,” would be if the “known
`
`inputs” referenced in Lemelson necessarily included “data of possible exterior
`
`objects and patterns of received waves from the possible exterior objects.”
`
`Further, it is not enough to merely show that Lemelson discloses a “trained
`
`pattern recognition algorithm” when there are numerous different ways to generate
`
`such an algorithm other than the manner required by the claims. The ‘057 patent
`
`claims do not just claim a “trained pattern recognition algorithm,” period. The
`
`added requirement that the algorithm be “generated from data of possible exterior
`
`objects and patterns of received waves from the possible exterior objects” must be
`
`also disclosed in the prior art for there to be anticipation.
`
`c.
`
`the claim
`inherently disclose
`Lemelson does not
`limitation—it could have involved generating the algorithm
`with simulated data
`
`Lemelson does not inherently disclose the claimed manner of generating a
`
`pattern recognition algorithm because there are several other ways that Lemelson
`
`could have generated its pattern recognition algorithm. First, the system in
`
`Lemelson could have been generated using simulated data, rather than data from
`
`
`
`14
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`possible exterior objects and patterns of received waves from the possible exterior
`
`objects. (See Exhibit 2001, Koutsougeras Decl. at ¶¶ 55-63.)
`
`Simulated data is data that does not include any “patterns of received waves
`
`from the possible exterior objects.” (Id.) Rather, it is generated by computer
`
`programs that simulate what sensors would be reading if they were detecting an
`
`object. (Id.) As Professor Koutsougeras explains, “[s]imulated data is therefore
`
`not data from objects or patterns of waves from objects—it is completely made-up
`
`data.” (Id. at ¶ 56.) In his declaration, he explains that as an analogy, simulated
`
`data is similar to a movie made with actors versus a cartoon. The cartoon would
`
`provide a rough approximation for what a person is expected to look like, but not
`
`nearly as accurate as a video with a real actor. (See id. at ¶ 57.)
`
`Professor Koutsougeras also explains that using simulated data for
`
`generating a pattern recognition algorithm for a vehicle could very well have been
`
`the “known inputs” referenced by Lemelson. (See id. at ¶¶ 55-63.) Lemelson
`
`claims priority to an application that was filed in 1993. (See id. at ¶ 58, citing
`
`Lemelson at cover.) In 1992, Pomerleau described in his thesis using simulated
`
`data to train a neural network on a vehicle.3 (See id.) In fact, Pomerleau devoted
`
`
`3 AVS again notes that the Pomerleau article was rejected as the basis for a ground
`
`for review. Pomerleau discloses two ways to train a neural network—with real
`
`
`
`
`
`15
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`an entire section of his thesis to what was titled “Training With Simulated Data.”
`
`(See id. at ¶ 58, citing Exhibit 2004, Pomerleau at p. 38.) Pomerleau explained that
`
`he believed at the time that “the only way to achieve variety in the training set
`
`sufficient to ensure that the network learns a general internal representation was to
`
`generate the training set synthetically.” (Id., citing Exhibit 2004 at p. 38.)
`
`Pomerleau explained that:
`
`To generate synthetic training data for the task of autonomous road
`
`following, I developed a program that generated aerial views of
`
`simulated stretches of roads and then used a model of the camera to
`
`back-project the aerial map into a 2D image of the road ahead. The
`
`
`data and with simulated data. The Board recognized that the Pomerleau article
`
`disclosed using real data only for generating a neural network for detecting a road
`
`(not objects). The Pomerleau article also vaguely refers to programming
`
`parameters of objects for the “simulated road generator,” but not for training with
`
`real data (nor, as AVS explained in its Preliminary Response, does the Pomerleau
`
`article provide an enabling disclosure even for simulated data of objects). (See
`
`Paper 17, AVS’s Prelim. Resp. at p. 39.) In any event, the Pomerleau Thesis (a
`
`different publication) was not asserted by Toyota in its Petition, and it is not the
`
`basis of any ground for review.
`
`
`
`16
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`simulated road image generator used nearly 200 parameters in order to
`
`generate a variety of realistic road images. Some of the most
`
`important parameters are listed in Figure 3.1.
`
`(Id. at ¶ 58, citing Exhibit 2004 at page 38.) And Pomerleau reported that
`
`generating the algorithm using simulated data worked. (See id. at ¶ 58, citing
`
`Exhibit 2004 at p. 40 (“Pomerleau reports in his thesis that, using the simulated
`
`training set of artificial road images, the network ‘could accurately drive Navlab I
`
`at a speed of 4 miles per hour along a 400 meter path through a wooded area of the
`
`CMU campus under sunny fall conditions.’”).)
`
`Professor Koutsougeras also discusses how the use of simulated data for
`
`training a neural network was widely known and used in other contexts as well.
`
`For example, he cites to U.S. Pat. No. 5,537,327, which involved the use of a
`
`trained neural network to detect impedance faults on a power line. (See Exhibit
`
`2001, Koutsougeras Decl. at ¶ 59.) That patent included claim 4 “wherein said
`
`neural network training is accomplished by the use of simulated data” and claim 6
`
`“wherein said neural network training is accomplished by applying actual data.”
`
`(See Exhibit 2003, U.S. Pat. No. 5,537,327 at claims 1, 4, and 6) (emphasis added).
`
`One reason why the “known inputs” of Lemelson may have been simulated
`
`data is because training with simulated data can be more desirable in some aspects
`
`compared to real-world data. Using simulated data has advantages in being able to
`
`
`
`17
`
`
`
`PATENT OWNER’S RESPONSE UNDER 37 CFR §42.120
`IPR2013-00419
`generate a large training set easily and ensure “balance” in a pattern recognition
`
`algorithm. (See, e.g., Exhibit 2004, Pomerleau Thesis at 38 (“I believed at the time
`
`that the only way to achieve variety in the training set sufficient to ensure that the
`
`network learns a general internal representation was to generate the training set
`
`synthetically.”).) (See also Exhibit 2001, Koutsougeras Decl. at ¶ 61-62.) For
`
`example, if a pattern recognition algorithm is generated with mostly compact
`
`sedans, and only a few SUV’s, it may tend to recognize only cars that are closer in
`
`appearance to a compact sedan. As another analogy to illustrate the importance of
`
`a large number of examples for generating an algorithm, take the example of a
`
`system for detecting if an individual is male or female. If all of the female
`
`examples have long hair and all of the male examples have short hair, the system
`
`may believe that a long-haired male is a female. Using simulated data ma