Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 1 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`UNITED STATES DISTRICT COURT
`SOUTHERN DISTRICT OF NEW YORK
`
`
`
`
`NETWORK-1 TECHNOLOGIES, INC.,
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`v.
`
`
`
`Plaintiff,
`
`GOOGLE LLC and YOUTUBE, LLC,
`
`
`
`
`
`
`
`
`Defendants.
`
`
`
`
`14 Civ. 2396 (PGG)
`
`14 Civ. 9558 (PGG)
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`PLAINTIFF NETWORK-1 TECHNOLOGIES, INC.’S
`BRIEF IN OPPOSITION TO DEFENDANTS’ MOTION FOR SUMMARY JUDGMENT
`
`
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 2 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`
`
`I.
`II.
`III.
`
`IV.
`V.
`V.
`
`TABLE OF CONTENTS
`
`The Siberia Version of Content ID Performs a Sublinear Approximate Nearest
`
`The overall search algorithm is a sublinear approximate nearest neighbor
`
`The overall search algorithm performs a search of reference extracted
`
`BOTH VERSIONS OF THE CONTENT ID SYSTEM PERFORM A SUBLINEAR
`APPROXIMATE NEAREST NEIGHBOR SEARCH OF REFERENCE EXTRACTED
`
`INTRODUCTION .............................................................................................................. 1
`BOTH VERSIONS OF CONTENT ID USE A SUBLINEAR SEARCH ......................... 2
`A.
`The Content ID LSH Version Uses a Sublinear Search ......................................... 3
`B.
`The Content ID Siberia Version Uses a Sublinear Search ...................................... 8
`FEATURES ...................................................................................................................... 12
`A.
`Neighbor Search of Reference Extracted Features ............................................... 13
`1.
`search ........................................................................................................ 15
`2.
`features ...................................................................................................... 17
`B.
`Neighbor Search of Reference Extracted Features ............................................... 19
`THE ASSERTED CLAIMS OF THE ’464 PATENT ARE NOT INVALID .................. 22
`PATENT ........................................................................................................................... 25
`CONCLUSION ................................................................................................................. 28
`
`The LSH Version of Content ID Performs a Sublinear Approximate Nearest
`
`AT LEAST A GENUINE ISSUE OF FACT EXISTS AS TO WHETHER GOOGLE’S
`CONTENT ID SYSTEM INFRINGES THE ASSERTED CLAIMS OF THE ’464
`
`i
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 3 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`TABLE OF AUTHORITIES
`
`CASES
`
`ACME Worldwide Enters, Inc. v. United States,
`146 Fed. Cl. 341 (Fed. Cir. 2019) ................................................................................................. 28
`
`Altiris, Inc. v. Symantec Corp.
`318 F.3d 1363 (Fed. Cir. 2003)................................................................................................ 25
`
`
`Anhydrides & Chemicals, Inc. v. United States,
`130 F.3d 1481 (Fed. Cir. 1997)................................................................................................. 26
`
`
`CBT Flint Partners v. Return Path, Inc.
`654 F.3d 1353 (Fed. Cir. 2011)................................................................................................. 25
`
`
`Conoco, Inc. v. Energy & Envtl. Int’l, L.C.,
`460 F.3d 1349 (Fed. Cir. 2006)................................................................................................. 19
`
`
`Dow Chem. Co. v. United States,
`226 F.3d 1334 (Fed. Cir. 2000)................................................................................................. 19
`
`
`Finisar Corp. v. DirectTV Group, Inc.,
`523 F.3d 1323 (Fed. Cir. 2008)................................................................................................. 26
`
`
`Hoffer v. Microsoft Corp.,
`405 F.3d 1326 (Fed. Cir. 2005)................................................................................................. 24
`
`
`In re Gabapentin Patent Litig.,
`503 F.3d 1254 (Fed. Cir. 2007)................................................................................................... 8
`
`
`Interactive Gift Exp., Inc. v. Compuserve Inc.
`256 F.3d 1323 (Fed. Cir. 2001)................................................................................................. 25
`
`
`Inventio AG v. ThyssenKrupp Elevator Ams. Corp.,
`No. 08-cv-874, 2013 WL 6627945 (D. Del. 2013) ................................................................... 24
`
`
`Lemelson v. General Mills, Inc.,
`968 F.2d 1202 (Fed. Cir. 1992)................................................................................................. 24
`
`
`Lockhart v. United States,
`136 S. Ct. 958 (2016) ................................................................................................................ 26
`
`
`Medtronic CoreValve, LLC v. Edwards Lifesci. Corp.,
`741 F. 3d 1359 (Fed. Cir. 2014)................................................................................................ 24
`
`
`
`ii
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 4 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`Simmons, Inc. v. Bombardier, Inc.,
`328 F. Supp. 2d 1188 (D. Utah 2004) ....................................................................................... 24
`
`
`Uniloc USA, Inc. v. Microsoft Corp.,
`632 F.3d 1292 (Fed. Cir. 2011)................................................................................................. 19
`
`
`Vitronics Corp. v. Conceptronic, Inc.,
`90 F.3d 1576 (Fed. Cir. 1996)................................................................................................... 28
`
`
`
`
`iii
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 5 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`I.
`
`INTRODUCTION
`
`Google’s motion for summary judgment addresses a number of discrete issues. For each,
`
`Google fails to establish its entitlement to summary judgment. Google relies on factually incorrect
`
`assertions, incorrect characterizations of the evidence, and legally incorrect premises to support its
`
`motion. Indeed, Google’s motion rests on the premise that 158 statements of fact are all each
`
`entirely undisputed. See Google’s Rule 56.1 Statement. Unsurprisingly, Google fails to identify
`
`such a sweeping panoply of undisputed facts, warranting denial of its motion at the outset. Further,
`
`the individual arguments put forward in Google’s motion each fails on its own terms.
`
`First, Google asserts that neither of the two versions of its accused Content ID system
`
`utilize a “sublinear” search consistent with the parties’ agreed construction of that term. Clear
`
`evidence shows that both versions use a sublinear search—as explained in detail by Network-1’s
`
`expert witness, Professor Mitzenmacher, and as admitted repeatedly in Google’s own documents
`
`and witness testimony. Genuine issues of fact clearly preclude summary judgment.
`
`Second, Google asserts that neither of the versions of the accused system utilize an
`
`approximate nearest neighbor search of reference extracted features. To make this argument,
`
`Google starts from the incorrect premise of examining only a subpart of the search algorithm
`
`identified by Network-1, rather than the entire algorithm that is actually accused. This incorrect
`
`focus alone is sufficient to reject Google’s motion on these issues. Google also fails to show that
`
`this issue can be decided on summary judgment because numerous genuine issues of material fact
`
`make summary judgment inappropriate. Professor Mitzenmacher, relying on Google’s own
`
`documents, computer code, and testimony, explains at length why each version of the accused
`
`system meets these claim elements.
`
`Third, Google asserts that a clear printing error that appears in one location on the ’464
`
`patent renders the patent invalid even though the correct version of the same information appears
`
`1
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 6 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`on the face of the patent itself, and in the original application for the patent. Specifically, Google
`
`argues that in the recitation of the priority chain of the patent on its cover, the patent refers to an
`
`application as a “continuation” when it was, in fact, a “continuation in part.” At column one of
`
`the patent, however, the application is correctly identified as a “continuation in part.” Google’s
`
`argument for invalidity on this basis is contradictory to clear Federal Circuit precedent that such
`
`printing errors do not invalidate the patent and that they are precisely the types of errors that the
`
`Court may readily correct.
`
`Finally, Google argues that it does not infringe the claims of the ’464 patent because it does
`
`not perform all of the claim steps in the order recited in the claim. Clear precedent from the Federal
`
`Circuit holds that claim steps need not be performed in the order written for infringement to be
`
`found. Google’s argument about how to interpret the language of the claim purports to rely on
`
`some unstated notion of grammar. In reality, Google ignores an actual rule of grammar, the last
`
`antecedent rule, firmly established in both Supreme Court and Federal Circuit precedent.
`
`Each of Google’s arguments fails to establish a basis for the Court to grant summary
`
`judgment. As such, Google’s motion should be denied in its entirety.
`
`II.
`
`BOTH VERSIONS OF CONTENT ID USE A SUBLINEAR SEARCH
`
`For both accused versions of Content ID (the Content ID LSH Version and the Content ID
`
`Siberia Version), the record is replete with evidence that the search used is “a search whose
`
`execution time scales with a less than linear relationship to the size of the data set to be searched,
`
`assuming computing power is held constant.” At a minimum, there are material factual disputes
`
`concerning the operation of both versions that are related to whether those versions meet the claim
`
`language, which precludes summary judgment concerning the sublinear claim limitation.
`
`2
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 7 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`A.
`
`The Content ID LSH Version Uses a Sublinear Search
`
`First, Google’s own documents and fact witness testimony indicate the search of the
`
`Content ID LSH Version was sublinear. For example:
`
`In general, [Content ID LSH Version] has no unscalable bottlenecks and hence is
`able to cope with increases in scale, with linear or sublinear growth in resource
`requirements depending on the resource. . . . Disk space will scale approximately
`linearly with growth in reference set size (but the amount is insignificant), other
`resources (CPU, RAM) will scale sublinearly.
`
`Ex. 40 at GOOG-NETWORK-00000418 (emphases added).1 Both CPU (central processing unit)
`
`and RAM (random access memory) resources are related to the amount of computing power, and
`
`therefore the execution time (if computing resources are held constant), needed to complete a
`
`search. Adding computing resources is functionally the same as adding time. Either you can add
`
`more computing resources and consume the same amount of time, or keep computing resources
`
`constant and add to the execution time. Thus, references to the scaling of computing resources are
`
`the same as references to the scaling of time if the computing resources are unchanged. See
`
`generally Ex. 27, Mitzenmacher Report at ¶¶ 192, 209-214. That these computing resources scale
`
`sublinearly with increasing data set size necessarily means that the execution time to perform a
`
`search by the Content ID LSH Version scales with a less than linear relationship to the size of the
`
`data to be searched. See id. ¶¶ 144, 209. The Content ID LSH version therefore is sublinear under
`
`the parties’ claim construction. Id.
`
`As another example, a paper co-authored by one of Google’s technical witnesses, Shumeet
`
`Baluja, states that “LSH [locality sensitive hashing] and other hash functions are sublinear in the
`
`number of elements examined compared to the size of the database.” Ex. 56 at GOOG-
`
`NETWORK-00613420 (emphasis added); Ex. 27, Mitzenmacher Report ¶¶ 144, 209. When asked
`
`
`1 As in Network-1’s Rule 56.1 Statement, exhibit references are to the Appendix of Exhibits, and
`references to Professor Mitzenmacher’s report include the materials cited in the referenced
`sections.
`3
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 8 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`about this statement, Dr. Baluja testified that the sublinearity discussed in this paper was “referring
`
`to the same sublinearly . . . discussed earlier” in his deposition. See Ex. 29, Baluja Depo. at 180:23-
`
`181:6. And earlier in his deposition, Dr. Baluja confirmed that he had the same understanding of
`
`“sublinear” as reflected in the parties’ claim construction:
`
`Q. Towards the end of the first paragraph, it says, “The system also provides
`good scaling characteristics. When the database in increased by 50 percent, we
`see that we have a sublinear computation increase while having no significant
`impact on recognition.” Do you see that?
`A. Yes, Yes.
`Q. What does that mean?
`A. So what that means is we will still consider all the elements in our repository
`without having to examine them in detail.
`Q. So using this LSH approach allows scaling of the system to be sublinear in
`that sense? Is that what you’re saying?
`A. Yes.
`
`Id. at 138:3-18 (testifying concerning Ex. 54 at GOOG-NETWORK-00610586) (emphasis
`
`added). Critically, Dr. Baluja was referring to sublinear scaling in reference to the amount of
`
`computing power/execution time required, just as the parties’ claim construction does.2 See Ex.
`
`27, Mitzenmacher Report ¶¶ 144, 209.
`
`Second, Network-1’s expert has testified unequivocally that the search algorithm of the
`
`Content ID LSH Version scales sublinearly. Dr. Mitzenmacher explained in detail that LSH tables
`
`or indexes use a data structure that is by design sublinear:
`
`145. Starting from the first step, the Content ID LSH Version is designed to
`determine a very small subset of the reference works in the database, in particular
`a sublinear subset, that could be possible matches to the input work being queried.
`This is through the creation of what is commonly referred to as an “inverted index”
`data structure, based on the LSH bands; only reference works that match in terms
`of the LSH bands are subject to further analysis. See, e.g., [Ex. 40 at] GOOG-
`NETWORK-00000412 (“The LSH (Locality Sensitive Hash) Tables are an
`
`
`2 In their brief, Defendants contend that Dr. Baluja was referring to some different (yet
`unspecified) meaning of sublinear. See Opening Br. at 24. However, examination of a more
`complete excerpt of Dr. Baluja’s testimony reveals that was not the case.
`4
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 9 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`inverted index of LSH bands (substrings extracted from the fingerprint to the IDs
`of videos containing that band.”).
`146. The inverted index is designed to be a sublinear data structure; that is, the
`inverted index, on a query, specifically does not go through each of the reference
`in the database to see if they match the query. Rather, in this setting, when given
`an LSH band, the inverted index can directly return a list of the reference works
`that that LSH bands, in time proportional to the number of matches. Hence the
`work done by the inverted index corresponds to the number of index hits, not the
`number of references. This is a general property of inverted indexes.
`
`Ex. 27, Mitzenmacher Report ¶¶ 145-146 (emphases added); see also id. ¶¶ 210-211. In short, Dr.
`
`Mitzenmacher explained that general properties of inverted indexes like that used in the Content
`
`ID LSH Version are that (a) they are sublinear data structures and (2) the work done (i.e., the
`
`computing power and therefore execution time required for a search) corresponds to the number
`
`of index hits or matches, rather than the number of references. See id. ¶¶ 145-146, 210-211.
`
`Faced with their own documents, testimony from their own technical witnesses, and
`
`testimony from Dr. Mitzenmacher, Defendants contend that “Network-1 has not presented any
`
`argument or evidence that the number of matches scales in a way that is less than proportional to
`
`the number of references to be searched.” Opening Br. at 23. But it logically follows from Dr.
`
`Mitzenmacher’s explanation (Ex. 27, Mitzenmacher Report ¶¶ 145-146, 210-211) that the number
`
`of matches scales in a sublinear (and not in a proportional or linear) fashion as the number of
`
`references to be searched increases. In fact, Dr. Mitzenmacher so testified at his deposition:
`
`Q. So just explain to me what you mean -- what does it mean to say that the Content
`ID LSH version system determines a sublinear subset [of matches]?
`A. So I think the point is that the work going in to like the number of things in the
`subset were sublinearly with the corresponding work or execution time to handle
`such objects, while also grow sublinearly in the setting of the context of claim
`construction.
`
`Ex. 34, Mitzenmacher Depo. at 169:15-23; id. at 180:24-181:3 (“[A]nd you’re saying the results
`
`returned is a sublinear function of the total set of reference works? A. Yes. That’s the -- by design
`
`of the inverted index.”); see also id. at 168:11-170:17, 172:4-173:3, 179:7-22, 183:23-184:11;
`
`276:6-22.
`
`5
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 10 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`Third, Defendants’ disagreement with Dr. Mitzenmacher is based on the experts’
`
`competing views of the functionality of certain LSH index lookup source code. Curiously,
`
`Defendants omit any discussion of this factual dispute from their opening brief.
`
`To understand this factual dispute, it helps to understand how the LSH bands of the
`
`reference videos are stored and indexed. These bands are indexed using a data structure known as
`
`BigTable. Ex. 28, Erb Depo. at 75:11-20; Network-1’s Response to Defendants’ Statement of
`
`Material Facts (“SMF”) ¶ 159. The parties agree that, “[i]n a conceptual sense, the LSH index
`
`could be visualized as a ‘table’ in which each unique LSH band is assigned its own ‘row’ and each
`
`reference video assigned its own ‘column.’” Defendants’ SMF ¶ 84; Network-1’s Response to
`
`Defendants’ SMF ¶ 84; see Opening Br. at 22; Ex. 34, Mitzenmacher Depo. at 212:21-213:16.
`
`The parties also agree that “[a] search of the LSH index using a particular query LSH band
`
`searched only the row assigned to that particular LSH band and did not search any of the other
`
`rows in the conceptual table.” Defendants’ SMF ¶ 85; Network-1’s Response to Defendants’ SMF
`
`¶ 85. But that is where the parties’ agreement on the LSH lookup process ends.
`
`In essence, the parties disagreement is rooted in whether a query based on an LSH band
`
`iterates over every reference video (represented by the conceptual columns) in the data set or not.
`
`Defendants’ expert Dr. Bhattacharjee opines that “[t]he search performed by the LSH Version of
`
`the Content ID system is not sublinear because
`
`
`
`84, Expert Report of Dr. Samrat Bhattacharjee (“Bhattacharjee Report”) ¶¶ 290-292. He further
`
`even if there are tens of millions of columns. Ex.
`
`elaborates that “[l]ine 488 [of the file
`
`thus demonstrates that for each LSH band lookup,
`
`
`
` The code
`
`
`
`Id. ¶ 294. He further states that this code interpretation is the basis for his opinion that the Content
`
`6
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 11 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`ID LSH Version is not sublinear and why he believes that the “execution time of the search
`
`increases proportionately with the number of reference videos used in connection with the Content
`
`ID system.” Id. ¶¶ 290-296.3
`
`Dr. Mitzenmacher disagrees with Dr. Bhattacharjee’s code interpretation of
`
` file, and rather opines that a query based on an LSH band
`
`
`
`
`
` the cells that correspond to reference videos that contain the given
`
`LSH band:
`
`Q. Okay. But you disagree that line – well, that the for loop that begins at line
`
`
`
`
`A. Correct.
`Q. Okay. And the basis for your disagreement is your general understanding of
`how BigTable works?
` right, so you’re
`A. That and the code itself, it’s describing
`pointing to the issue of the -- you’re pointing to the line code. If it -- if it -- this for
`loop if it was just going through the table sequentially, like the typical way you’d
`write that is from I equals zero [to], I equals, you know, the biggest entry in the
`table, I plus plus, right, you would just literally do a sequential walkthrough the
`cells, right?
`
`
`Q. That’s -- that’s
`A. I -- I agree
`
`
`
`
`
`
`Ex. 34, Mitzenmacher Depo. at 222:13-223:11 (emphases added); see also id. at 223:12-227:22;
`
`Network-1’s Response to Defendants’ SMF ¶¶ 160-162; Ex. 33, Bhattacharjee Depo. at 120:16-
`
`20 (admitting that Content ID LSH Version Big Table does not store empty cells); id. at 105:15-
`
`20 (admitting that the rows in the Content ID LSH Version Big Table vary in size).
`
`
`3 Network-1 intends to file a motion in limine to preclude Dr. Bhattacharjee and Defendants from
`relying on this source code file because this file was not produced during the fact discovery period.
`Regardless, Network-1 details its (and its expert’s) disagreement with Dr. Bhattacharjee’s analysis
`here.
`7
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 12 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`In sum, Defendants’ motion for partial summary judgment concerning the sublinearity of
`
`the Content ID LSH Version should be denied for at least three reasons. First, Defendants own
`
`documentation concerning the Content ID LSH Version, as well as Dr. Baluja’s testimony, indicate
`
`the system’s search was “sublinear,” as that term is defined in the parties’ agreed-upon claim
`
`construction. Second, Network-1’s expert provided detailed explanations of why the inverted
`
`index data structure used in the Content ID LSH Version results in a search algorithm that scales
`
`sublinearly. Third, because the parties’ experts dispute how a particular piece of code operates,
`
`and that factual dispute goes to the heart of whether Content ID LSH Version uses a sublinear
`
`search, summary judgment should be denied for this additional reason. See, e.g., In re Gabapentin
`
`Patent Litig., 503 F.3d 1254, 1259-62 (Fed. Cir. 2007) (noting that whether a claim limitation is
`
`met by an accused product is a question of fact, and reversing grant of summary judgment of non-
`
`infringement because competing expert testimony created genuine factual disputes concerning
`
`whether a certain claim limitation was met).
`
`B.
`
`The Content ID Siberia Version Uses a Sublinear Search
`
`As a preliminary matter, Defendants contend that Network-1’s infringement theory for the
`
`Content ID Siberia Version is not legally cognizable because it requires that Defendants alter or
`
`reprogram the accused system before it can infringe Network-1’s patent. See Opening Br. at 19-
`
`20. Defendants misapprehend Network-1’s infringement allegations. Network-1 is not alleging
`
`that the Content ID Siberia Version must be “reprogrammed” in order to be infringing; rather,
`
`Network-1 is alleging that the system is infringing because its search algorithm is “sublinear,” i.e.,
`
`“a search whose execution time scales with a less than linear relationship to the size of the data set
`
`to be searched, assuming computing power is held constant.” Network-1’s expert Dr.
`
`Mitzenmacher explained that he is likewise not proposing alterations or changes to the system to
`
`somehow make it infringing, but rather “that the algorithm itself is sublinear because it has this
`
`8
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 13 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`ability to change and adapt to these situations” of growing data set size. Ex. 34, Mitzenmacher
`
`Depo. at 128:19-129:22 (emphasis added). Indeed, during the design phase of the Content ID,
`
`Defendants recognized that the Content ID Siberia Version would need to use a sublinear search
`
`to be practicable. Ex. 85 at GOOG-NETWORK-00704247-50 (“Given the size of the dataset, it
`
`is clear that we will need a (customized) distribute searching service and a sublinear search with
`
`hashed representation . . . .”) (emphasis added).
`
`Like the Content ID LSH Version, the search algorithm of the Content ID Siberia Version
`
`begins with a sublinear index lookup step, which for Siberia is known as ScaM or “scalable
`
`matching.” Ex. 30, Pasula Depo. at 52:12-62:18, 67:7-70:16. Rather than LSH bands, Siberia’s
`
`reference indices
`
` Id. at 45:7-
`
`46:20; Ex. 35, Kumar Depo. at 18:4-29:19. For a given reference video, the Content ID Siberia
`
`Version generates one or more sequences of embeddings, and each of the embeddings is then
`
` Ex. 30, Pasula Depo. at 45:7-46:20; Ex. 35, Kumar Depo. at 18:4-29:19.
`
`
`
`
`
`
`
` Ex. 30, Pasula Depo. at 39:11-44:17.
`
`With that basic understanding of Siberia’s indexing, it is evident that Defendants own
`
`documents indicate that as the size of the reference index increases, the search algorithm of Content
`
`ID Siberia Version is designed to scale sublinearly:
`
`9
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 14 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`
`
`Ex. 67 at GOOG-NETWORK-00704057 (emphasis added). Here, the total amount of computing
`
`power or execution time (“Query Cost (CPU)”) is reflected on the y-axis, and the size of the dataset
`
`is reflected on the x-axis. Id.; Ex. 30, Pasula Depo. at 77:19-78:17. The graph above shows that
`
`the
`
` used in the Content ID Siberia Version has a Query Cost
`
`(amount of computing power required) that increases in a sublinear fashion in relation to the size
`
`of the dataset, or “Dataset Scale.” Ex. 27, Mitzenmacher Report ¶¶ 152-156; see also Ex. 35,
`
`Kumar Depo. at 157:6-15 (“Content ID uses a very specific type of t
`
`
`
` It is just a flat partitioning of data.”). The graph above does not reflect
`
`alterations or reprogramming of the Content ID Siberia Version, but rather simply that the Query
`
`Cost or computing power required scales sublinearly with increasing Dataset Scale. Ex. 27,
`
`Mitzenmacher Report ¶¶ 152-156; see also Ex. 34, Mitzenmacher Depo. at 131:15-133:18 (further
`
`discussing how and why increasing the number of shards can result in sublinear scaling of Query
`
`Cost); Ex. 69 at GOOG-NETWORK-00717464 (“ScaM is handling a number of neighbor search
`
`problems like Content ID copyright detection. This is currently done by using sub-linear
`
`approximate search like hashing-based and tree-based solutions with CPUs.”) (emphasis added).
`
`10
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 15 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`In addition, even though
`
`“has query-costs that grow sub-linearly for increases in
`
`database size” (Ex. 67 at GOOG-NETWORK-00704057 (emphasis added)), Defendants
`
`themselves have recognized that there are various parameters in the search algorithm that can be
`
`adjusted as the size of the dataset increases to further enhance the system’s efficiency. For
`
`example, according to Defendants, one such “tunable knob” is
`
`
`
`Ex. 64 at GOOG-NETWORK-00702308 (listing
`
` as a “Tunable
`
`knob” in Siberia’s search). Dr. Mitzenmacher provided a detailed explanation of why increasing
`
` as the data set size increases, even if the
`
`remains
`
`unchanged, results in sublinear scaling. Ex. 27, Mitzenmacher Report ¶¶ 157-160; Ex. 34,
`
`Mitzenmacher Depo. at 137:6-142:21. This is merely an example of how and why “the algorithm
`
`itself is sublinear because it has this ability to change and adapt to these situations.” Ex. 34,
`
`Mitzenmacher Depo. at 128:19-129:22.
`
`Defendants contend that “the execution time of the search would continue to scale linearly
`
`with the size of the data set of be searched
`
` because the
`
`system would continue to search the same fixed fraction data set.” Opening Br. at 21. Defendants
`
`contend this is so because “source code dictates that the system will
`
`
`
` Id. In support of this statement concerning
`
`source code, Defendants point to paragraph 302 of Dr. Bhattacharjee’s expert report, but that
`
`paragraph contains no analysis or mention of source code to support this “fixed fraction” source
`
`code theory. See Ex. 84, Bhattacharjee Report ¶ 302. Rather,
`
` appears to be
`
`based on
`
`certain implementation of the video reference
`
`index, not on any hard-coded “fixed fraction.” See Opening Br. at 21 n.5; Defendants’ SMF ¶ 37,
`
`50. If
`
`“tunable knob” were adjusted without any other changes,
`
`
`
`11
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 16 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`would be examined than that proposed by Defendants. See Ex. 64 at
`
`GOOG-NETWORK-00702308.
`
`Finally, Network-1 agrees that “if additional references were added to the existing
`
`shard/partition structure, the [index lookup] portion of the search would scale linearly.” Opening
`
`Br. at 16 (quoting Ex. 27, Mitzenmacher Report ¶ 229). However, this is not the way in which the
`
`search algorithm Content ID Siberia Version was designed to be used, as demonstrated by
`
`Google’s own Siberia documentation.
`
`Defendants’ motion for partial summary judgment concerning the sublinearity of the
`
`Content ID Siberia Version should likewise be denied for the reasons detailed above. Defendants
`
`own documentation concerning the Content ID Siberia Version indicate the system’s search is
`
`“sublinear” under the parties’ agreed-upon claim construction. In addition, Network-1’s expert
`
`explained that Siberia’s search algorithm itself is sublinear because it is designed to adapt to
`
`maintain efficiency when faced with growing data set size. At a minimum, there are material facts
`
`in dispute concerning the design and operation of the system that are key to determining whether
`
`this claim limitation is met.
`
`III. BOTH VERSIONS OF THE CONTENT ID SYSTEM PERFORM A SUBLINEAR
`APPROXIMATE NEAREST NEIGHBOR SEARCH OF REFERENCE
`EXTRACTED FEATURES
`
`In Section III of its motion, Defendants argue sequentially that neither of the two versions
`
`of the Content ID system perform a sublinear approximate nearest neighbor search of reference
`
`extracted features. This refers to elements of claim 33 of the ’237 patent and claims 34 and 35 that
`
`depend from claim 33. In particular, Defendants refers to element 33(b): “determining, by the
`
`computer system, an identification of the media work using the media work extracted features to
`
`perform a sublinear approximate nearest neighbor search of reference extracted features of
`
`reference identified media works.” Both versions of Defendants’ accused Content ID system
`
`12
`
`

`

`Case 1:14-cv-02396-PGG-SN Document 241 Filed 11/12/20 Page 17 of 34
`CONFIDENTIAL OUTSIDE COUNSEL ONLY
`PROSECUTION/ACQUISITION BAR MATERIALS
`
`perform this claim element. As to each version, Defendants attempt to focus only on limited steps
`
`within the search algorithm, rather than on the entirety of the search algorithm that Network 1
`
`accuses. Essentially Defendants argue that some sub-portion of its system, standing alone, does
`
`not infringe. Defendants’ arguments do not even focus on the actual accused system. This would
`
`be akin to arguing that a car could not meet a claim requirement for having an internal combustion
`
`engine because the carburetor by itself was not an internal combustion engine. The argument is
`
`irrelevant and meaningless. Defendants’ erroneo