`__________________
`
`BEFORE THE PATENT TRIAL AND APPEAL BOARD
`__________________________________________________________________
`
`TOYOTA MOTOR CORPORATION
`
`Petitioner
`
`v.
`
`AMERICAN VEHICULAR SCIENCES,
`
`Patent Owner
`
`Patent No. 6,772,057
`Issue Date: Aug. 3, 2004
`Title: VEHICULAR MONITORING SYSTEMS
`__________________________________________________________________
`
`REPLY DECLARATION OF NIKOLAOS PAPANIKOLOPOULOS, PH.D.
`
`
`Case No. IPR2013-00419
`__________________________________________________________________
`
`IPR2013-00419 - Ex. 1023
`Toyota Motor Corp., Petitioner
`1
`
`
`
`I, Nikolaos Papanikolopoulos, Ph.D., hereby further declare and state as follows:
`
`I.
`BACKGROUND
`1. My employment and compensation information have not changed since I
`
`submitted my original declaration in support of Toyota’s Petition for Inter Partes
`
`Review of U.S. Patent No. 6,772,057 (“the ’057 patent”).
`
`2.
`
`A copy of my updated curriculum vitae is included herewith.
`
`II. ASSIGNMENT AND COMPENSATION
`3.
`I submit this declaration in support of Toyota’s Reply to Patent Owner’s
`
`Response (Paper 33, hereinafter “Response”) and in response to the Declaration
`
`(Exhibit 2001) and Deposition Testimony (Exhibit 1022) of Cris Koutsougeras.
`
`4.
`
`Specifically, I have been asked to respond to Dr. Koutsougeras’s opinions
`
`regarding the disclosure in U.S. Patent No. 6,553,130 (“Lemelson”) relating to neural
`
`network training.
`
`5.
`
`The opinions expressed in this declaration are not exhaustive of my opinions
`
`on the patentability of any of the claims in the ’057 patent. Therefore, the fact that I
`
`do not address a particular point should not be understood to indicate any agreement
`
`on my part that any claim otherwise complies with the patentability requirements.
`
`In forming my opinion I have reviewed the following additional sources:
`
`
`•
`
`Declaration of Chris Koutsougeras, PhD in Support of AVS’s Response
`
`Under 37 CFR §42.120 (Ex. 2001).
`
`•
`
`Decision on Institution of Inter Partes Review for U.S. Patent No.
`
`
`
`1
`
`
`
` 2
`
`
`
`•
`
`•
`
`•
`
`•
`
`6,772,057 (Paper 19).
`
`Patent Owner’s Response (Paper 33).
`
`U.S. Patent No. 5,537,327 (Exhibit 2003).
`
`Pomerleau, “Neural Network Perception for Mobile Robot Guidance”
`
`(Exhibit 2004).1
`
`The transcript from the deposition of Dr. Cris Koutsougeras in
`
`connection with this case (Exhibit 1022).
`
`6.
`
`The opinions expressed in this declaration are my personal opinions and do not
`
`reflect the views of University of Minnesota.
`
`III. ANALYSIS
`A.
`Preliminary Understanding of Dr. Koutsougeras’ Positions
`As a preliminary matter, I understand from Dr. Koutsougeras’s declaration and
`
`7.
`
`deposition that he divided the data that could have been used for pattern recognition
`
`algorithm training (in 1995) into three areas: training with data and waves from actual
`
`objects (“real data”), training with simulated data and waves (“simulated data”), and
`
`training with “data and waves not representing exterior objects to be detected”
`
`(“partial data”). Ex. 1022 at 86:25-87:14, 132:24-138:5, 163:18-164:7. As I understand
`
`
`1
`I was previously familiar with this paper in the context of my work on
`
`IPR2013-00424 and from my interaction with Dean Pomerleau and the NavLab
`
`vehicle.
`
`
`
`2
`
`
`
` 3
`
`
`
`it, Dr. Koutsougeras opined that only training with real data would meet the claim
`
`limitation “pattern recognition algorithm generated from data of possible exterior
`
`objects and patterns of received waves.”
`
`8.
`
`As I understand it, Dr. Koutsougeras further opined that Lemelson’s disclosure
`
`of training with “known inputs” does not necessarily mean training with “real data”
`
`because it could have been referring instead to “simulated data” or “partial data.”
`
`9.
`
`For the reasons I discuss below, I disagree with Dr. Koutsougeras’s
`
`interpretation of Lemelson. In my opinion, one of ordinary skill in the art would have
`
`understood the phrase “known inputs” in Lemelson to refer to “real data” because
`
`Lemelson’s neural network was trained to identify exterior objects, and one of
`
`ordinary skill in 1995 would have known that training with “real data” would have
`
`yielded the best results for this purpose. One of ordinary skill in the art would not
`
`have understood that “known inputs” referred to simulated data or partial data in the
`
`context of Lemelson’s disclosure, since one of ordinary skill would not have had any
`
`reason to believe that those categories of data would have been effective for the
`
`purpose of object identification (e.g., for the purpose of identification of a pedestrian).
`
`B. One of Ordinary Skill Would Have Understood that Training of
`the Lemelson Neural Network Would Have Used Real Data
`10. Lemelson discloses a collision avoidance system, wherein a neural network is
`
`used to identify many different types of objects that could present themselves as
`
`hazards on a roadway, including, for example, road barriers, trucks, automobiles,
`
`
`
`3
`
`
`
` 4
`
`
`
`pedestrians, signs and symbols, etc. Ex. 1002 at 5:41-43; 8:1-6. Lemelson explains:
`
`Neural networks used in the vehicle . . . warning system are trained to
`
`recognize roadway hazards which the vehicle is approaching including
`
`automobiles, trucks, and pedestrians. Training involves providing
`
`known inputs to the network resulting in desired output responses. The
`
`weights are automatically adjusted based on error signal measurements
`
`until the desired outputs are generated. Various learning algorithms may
`
`be applied. Adaptive operation is also possible with on-line adjustment
`
`of network weights to meet imaging requirements.
`
`Ex. 1002 at 8:1-6.
`
`11. One of ordinary skill in the art would have understood the phrase “known
`
`inputs,” and would have understood that it referred to the use of real sensor data in
`
`the context of Lemelson. For example, one of ordinary skill would have understood
`
`that training a neural network could involve putting actual examples of real-world
`
`objects in front of a camera, imaging them, and providing feedback to the neural
`
`network as to the desired output responses corresponding to those images.
`
`12. As set forth below, it is my opinion that one of ordinary skill would not have
`
`understood the phrase “known inputs” in the context of Lemelson to refer to “partial
`
`data” or “simulated data” because one of ordinary skill would have recognized that
`
`neither of these categories would have been effective for the intended purpose of
`
`training a neural network to identify various types of exterior objects. I often refer to
`4
`
`
`
`
`
` 5
`
`
`
`these ineffective training routines to my students as “garbage in–garbage out.”
`
`C. Dr. Koutsougeras’ Description of Partial Data is Inaccurate
`13. One of ordinary skill in the art would not have understood the phrase “known
`
`inputs” to refer to partial data, such as from license plates, tail lights or rear windows
`
`because they would have recognized that training a neural network with partial data
`
`would not have been successful for the purpose of identifying different exterior
`
`objects.
`
`14.
`
`For example, one of ordinary skill in the art would not have been able to use
`
`partial data, such as rear windows, license plates, or taillights to identify, classify, or
`
`locate pedestrians and to distinguish them from vehicles. While partial data may be
`
`useful in certain isolated situations, such as when there is only a single object of
`
`interest, partial data is not useful when there are many possible objects that need to be
`
`identified, such as is the case in Lemelson.
`
`15. Also, the presence of occlusions and/or shadows in the environment exterior
`
`to a vehicle complicates training since these occlusions and shadows may completely
`
`mask partial data.
`
`16. One of ordinary skill in the art would not have understood that a neural
`
`network used for exterior object identification would be trained with partial data.
`
`Rather, one of ordinary skill would have understood that a neural network would be
`
`trained with all available sensor information to associate particular sensor information
`
`with desired output responses. The purpose of training a neural network is to identify
`
`
`
`5
`
`
`
` 6
`
`
`
`the particular features in a scene that are important and that are indicative of the
`
`exterior object of interest. On the other hand, to detect objects using partial data, one
`
`of ordinary skill would already expect to know the particular features that are
`
`important and indicative of the exterior object of interest (e.g., the outline and corner
`
`points of a license plate). None of the examples that Dr. Koutsougeras points to in
`
`his declaration involved training a neural network with partial data prior to 1995. It is
`
`also worth noting that Dr. Koutsougeras points to a pattern recognition system I
`
`worked on to detect license plates. Contrary to Dr. Koutsougeras’s assumption, this
`
`system did not involve trained pattern recognition.
`
`17. Accordingly, one of ordinary skill would not have expected that training the
`
`neural network of Lemelson with partial data would have resulted in a system capable
`
`of identifying the exterior objects required by Lemelson, such as “pedestrians, barriers
`
`and dividers, turns in the road, signs and symbols.” Ex. 1002 at 5:42-43.
`
`D. Dr. Koutsougeras’ Description of Simulated Data and His
`Characterizations of Pomerleau are Inaccurate
`18. One of ordinary skill in the art would not have understood the phrase “known
`
`inputs” to refer to simulated data because they would have recognized that training a
`
`neural network with simulated data would not have been successful for the purpose of
`
`identifying different exterior objects. One of ordinary skill in the art in 1995 would
`
`have known that the generation of simulated data was not sophisticated enough to
`
`allow for training the type of neural network described by Lemelson. Thus, even if
`
`
`
`6
`
`
`
` 7
`
`
`
`simulated data were used, the result would have likely been “garbage in–garbage out.2”
`
`19. One of ordinary skill in the art would have known that simulated data suffered
`
`from several problems in 1995.
`
`20.
`
`First, generation of simulated data would have required a lot of computer
`
`power and special equipment, neither of which were disclosed by Lemelson. See, e.g.,
`
`the Warp supercomputer used by Pomerleau, Ex. 2004 at 40. Lemelson does not
`
`disclose any computer hardware or methods for generating simulated data. See, e.g.,
`
`1002 at Fig. 1.
`
`21.
`
`Second, the Lemelson neural network was trained to identify “other vehicles,
`
`pedestrians, barriers and dividers, turns in the road, signs and symbols.” As of 1995,
`
`
`2. I note that AVS has taken two quotes from my deposition regarding simulated data
`
`out of context. First, the system I was talking about on page 48 of the transcript was
`
`my own system for identifying only license plates. See Ex. 1025 at 41:4-49:21. This
`
`system was not a collision avoidance system where different types of exterior objects
`
`needed to be identified. I was only locating license plates. Second, the question on
`
`page 102 referred to a system from the present day and not from 1995. See Ex. 1025
`
`at 102:18-22. Both of my statements at my deposition are fully consistent with my
`
`opinion here that one of ordinary skill in the art at the time of the publication of
`
`Lemelson would only have expected to use real data as a known input to the neural
`
`network.
`
`
`
`7
`
`
`
` 8
`
`
`
`one of ordinary skill in the art would not have expected that a simulated data set could
`
`be readily generated that could accurately represent all exterior objects described by
`
`Lemelson as perceived by sensors on a vehicle. This type of simulation would have
`
`required modeling of both a moving camera and moving objects in a scene, such as
`
`pedestrians, which would have resulted in a very complex data set. Furthermore, one
`
`of ordinary skill would have recognized that all of these complexities would have been
`
`obviated by simply training the system with real data in a variety of situations.
`
`22. Dr. Koutsougeras cites to a document that I discussed at length in a declaration
`
`I submitted in connection with IPR2013-00424: Dean Pomerleau’s 1992 Thesis. Ex.
`
`2004. Dr. Koutsougeras relies on the disclosure of Pomerleau’s Thesis as proof that
`
`one of ordinary skill would have used simulated data in the Lemelson system.
`
`However, Dr. Koutsougeras fails to read and comprehend the entirety of Pomerleau’s
`
`disclosures in several ways.
`
`23.
`
`First, the Board determined that a similar Pomerleau reference (Ex. 1008) only
`
`attempts to detect the road surface and does not identify “exterior objects,” as that
`
`term was construed. Paper 19, Decision on Institution at 12-14, 35-36. Pomerleau
`
`does not stand for the proposition that one of ordinary skill in the art would have
`
`thought one could use simulated data to train the system of Lemelson to identify
`
`exterior objects.
`
`24.
`
`Second, the Pomerleau thesis concluded that training with simulated data “has
`
`serious drawbacks.” Ex. 2004 at p. 40. Ultimately, Pomerleau concluded that
`8
`
`
`
`
`
` 9
`
`
`
`simulated data should not be used to train a system. Id. at pp. 40, 56. He reached this
`
`conclusion despite the fact that Pomerleau’s computational needs were much less
`
`demanding than required by Lemelson. Pomerleau explained that “differences
`
`between the synthetic road images on which the network was trained and the real
`
`situations on which the network was tested often resulted in poor performance in real
`
`driving situations.” Ex. 2004 at p. 40. Pomerleau further stated: “[W]hile relatively
`
`effective at training the network to drive under the limited conditions of a single-lane
`
`road, it quickly became apparent that extending the synthetic training paradigm to deal
`
`with more complex situations such as multi-lane and off-road driving would require
`
`prohibitively complex training data generators.” Id. Because of these drawbacks,
`
`Pomerleau concluded that, “[g]enerating realistic artificial training data proved
`
`impractical for all but the simplest driving situations.” Ex. 2004 at p. 56.
`
`25. One of ordinary skill in the art would not have understood that the phrase
`
`“known inputs” in Lemelson referred to “simulated data.”
`
`26. Dr. Koutsougeras also cites to U.S. Patent No. 5,537,327, in his discussion of
`
`simulated training data. However, his citation to the ’327 patent is misplaced. The
`
`disclosure of the ’327 patent relates to a different subject matter from that disclosed
`
`by the ’057 patent or the Lemelson reference. It relates to the use of neural networks
`
`to identify fault impedances in electrical power systems. The ’327 patent does not
`
`involve classification, identification, or location of objects exterior to a vehicle, or for
`
`that matter, any exterior monitoring from a vehicle at all. Furthermore, the ’327
`9
`
`
`
`
`
`10
`
`
`
`patent accounts for none of the complications that would have arisen when
`
`identifying possible exterior objects that could collide with a vehicle, as in Lemelson.
`
`Accordingly, the ’327 patent would not have indicated to one of ordinary skill in the
`
`art that “simulated data” could have been used as a known input to the Lemelson
`
`system.
`
`
`
`Date: May 26, 2014
`
`
`
`______________________________
`Nikolaos Papanikolopoulos, Ph.D.
`
`
`
`10
`
`
`
`11
`
`