`
`Exhibit 35
`(Partially Redacted)
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 2 of 201
`
` 1 UNITED STATES DISTRICT COURT
`
` 2 SOUTHERN DISTRICT OF NEW YORK
`
`
`
` 3 CIVIL ACTION NO. 1:14-cv-02396-PGG
`
`
` 4
`
` 5 - - - - - - - - - - - - - - - - - - - - - - - -
`
` 6 NETWORK-1 TECHNOLOGIES, INC.,
`
` 7 Plaintiff,
`
` 8 vs.
`
` 9 GOOGLE, INC., and YouTube,
`
`10 Defendants.
`
`11 - - - - - - - - - - - - - - - - - - - - - - - -
`
`12
`
`13 PROSECUTION/ACQUISITION BAR MATERIALS
`
`14 CONFIDENTIAL OUTSIDE ATTORNEYS ONLY
`
`15 HIGHLY CONFIDENTIAL SOURCE CODE
`
`16 TRANSCRIPT OF THE STENOGRAPHIC NOTES OF
`
`17 THE DEPOSITION OF SANJIV KUMAR
`
`18 October 15, 2019
`
`19
`
`20
`
`21
`
`22
`
`23
`
`24 LORRAINE B. ABATE, Notary Public
` 456335
`25
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 3 of 201
`
` 1 UNITED STATES DISTRICT COURT
`
` 2 SOUTHERN DISTRICT OF NEW YORK
`
` 3 CIVIL ACTION NO. 1:14-cv-02396-PGG
`
` 4
`
` 5 - - - - - - - - - - - - - - - - - - - - - - - -
`
` 6 NETWORK-1 TECHNOLOGIES, INC.,
`
` 7 Plaintiff,
`
` 8 vs.
`
` 9 GOOGLE, INC., and YouTube,
`
`10 Defendants.
`
`11 - - - - - - - - - - - - - - - - - - - - - - - -
`
`12
`
`13 PROSECUTION/ACQUISITION BAR MATERIALS
`
`14 CONFIDENTIAL OUTSIDE ATTORNEYS ONLY
`
`15 HIGHLY CONFIDENTIAL SOURCE CODE
`
`16 TRANSCRIPT of the stenographic notes of
`
`17 the deposition of SANJIV KUMAR in the above-entitled
`
`18 matter, as taken by and before LORRAINE B. ABATE, a
`
`19 Certified Shorthand Reporter and Notary Public of the
`
`20 State of New York and Registered Professional
`
`21 Reporter, held at the offices of Williams & Connolly,
`
`22 LLP, 650 Fifth Avenue, New York, New York, on October
`
`23 15, 2019, commencing at 8:59 a.m., pursuant to
`
`24 Notice.
`
`25
`
`1
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 4 of 201
`
` 1
`
` 2 A P P E A R A N C E S:
`
` 3
`
` 4 RUSS AUGUST & KABAT, ESQS.
`
` 5 Attorneys for the Plaintiff
`
` 6 12424 Wilshire Boulevard, 12th Floor
`
` 7 Los Angeles, California 90025
`
` 8 BY: BRIAN LEDAHL, ESQ.
`
` 9 MARC FENSTER, ESQ.
`
`10 (310)826-7474
`
`11 bledahl@raklaw.com
`
`12 mfenster@raklaw.com
`
`13
`
`14 WILLIAMS & CONNOLLY, LLP
`
`15 Attorneys for the Witness
`
`16 725 Twelfth Street NW
`
`17 Washington DC 20005
`
`18 BY: GRAHAM SAFTY, ESQ.
`
`19 SAMUEL BRYANT DAVIDOFF, ESQ.
`
`20 (202) 434-5548
`
`21 gsafty@wc.com
`
`22 sdavidoff@wc.com
`
`23 A L S O P R E S E N T:
`
`24 James Brady, Videographer
`
`25 Demarron Berkley, Esq. Google In-House Counsel
`
`2
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 5 of 201
`
` 1
`
` 2 I N D E X
`
` 3
`
` 4 WITNESS EXAMINATION BY PAGE
`
` 5 Sanjiv Kumar Mr. Ledahl 4, 174
`
` 6 Mr. Safty 168
`
` 7
`
` 8 E X H I B I T S
`
` 9 EXHIBIT PAGE
`
`10 1 August 2015 Quantization Based Inner
`
`11 Product Search (QUIPS)with LUT 16
`
`12 Option Presentation 66
`
`13 2 Siberia Primer Description 76
`
`14 3 ScaM: Scalable Matching, A Research
`
`15 Perspective Presentation 81
`
`16 4 ScaM 2.0 Documents 94
`
`17 5 ScaM:Scalable Matching Presentation 103
`
`18 6 Content ID Siberia A ScaM
`
`19 Perspective Presentation 111
`
`20 7 ScaM: Scalable Matching Technical
`
`21 Deep Dive Presentation 154
`
`22 8 Amended Joint Claim Construction Chart 169
`
`23 DIRECTIONS NOT TO ANSWER
`
`24 PAGE
`
`25 174-176
`
`3
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 6 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 THE VIDEOGRAPHER: Today's date is
`
` 3 October 15th, 2019. The time is 8:59 a.m. My
`
` 4 name is Jim Brady. I'm the videographer. We're
`
` 5 here today at Williams & Connolly, 650 5th
`
` 6 Avenue, New York, New York. We're here today in
`
` 7 the matter of Network-1 Technology versus
`
` 8 Google. Today's witness' name is Sanjiv Kumar.
`
` 9 May I ask now that the attorneys please
`
`10 introduce themselves and for the court reporter
`
`11 to swear in the witness.
`
`12 MR. LEDAHL: Brian Ledahl from Russ
`
`13 August & Kabat on behalf of the plaintiff.
`
`14 MR. SAFTY: Graham Safty from Williams &
`
`15 Connolly on behalf of defendants Google and
`
`16 YouTube. With me is Sam Davidoff, also from
`
`17 Williams & Connolly on behalf of defendants.
`
`18 S A N J I V K U M A R,
`
`19 Having been first duly sworn by a Notary
`
`20 Public of the State of New York, was
`
`21 examined and testified as follows:
`
`22 EXAMINATION BY MR. LEDAHL:
`
`23 Q. Good morning, Mr. Kumar.
`
`24 A. Good morning.
`
`25 Q. Have you ever had your deposition taken
`
`4
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 7 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 before?
`
` 3 A. No.
`
` 4 Q. Okay. I'm going to go over a couple of
`
` 5 ground rules, if you will, to hopefully make the
`
` 6 process just a little bit smoother as we go through
`
` 7 the day.
`
` 8 First, as you probably noted already, we
`
` 9 have a court reporter who takes down a transcript of
`
`10 all of the things that we say during the deposition.
`
`11 There are a couple of things I'm going to ask you to
`
`12 try to remember to do to make that transcript as
`
`13 clear and as accurate as possible.
`
`14 The first is if you could wait until I
`
`15 finish my questions to begin answering, and
`
`16 similarly, I'll try to wait until you finish your
`
`17 answers to begin asking my next question. That way,
`
`18 only one of us is talking at a time, because it's
`
`19 very difficult for the court reporter to take down
`
`20 more than one of us at a time.
`
`21 Do you understand?
`
`22 A. Yes.
`
`23 Q. The second is that because nods of the
`
`24 head or phrases like um-hum and uh-uh are extremely
`
`25 difficult to accurately capture on a transcript, I'm
`
`5
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 8 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 going to ask you to do your best to remember to say
`
` 3 yes or no, as opposed to using those expressions. I
`
` 4 may try to remind you if I notice it, but it's not me
`
` 5 being rude. I just want to make sure we have an
`
` 6 accurate record.
`
` 7 Do you understand?
`
` 8 A. Yes, I do.
`
` 9 Q. Okay. At any point today, if you don't
`
`10 understand my question, please let me know. I'll do
`
`11 my best to clarify. Otherwise, if you answer, I'll
`
`12 assume you understood the question. Fair enough?
`
`13 A. Yes.
`
`14 Q. Do you understand that the oath you took
`
`15 a few moments ago is the same oath you would take as
`
`16 if you were testifying in court in front of a judge
`
`17 and a jury?
`
`18 A. Yes, I do.
`
`19 Q. And you understand that oath carries the
`
`20 same penalties of purgery as if you were testifying
`
`21 in court?
`
`22 A. Yes, I do.
`
`23 Q. Are you on any medication, or is there
`
`24 any other health reason that would prevent you from
`
`25 giving your best and most accurate testimony today?
`
`6
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 9 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 A. No.
`
` 3 Q. Okay. As we go along today, we'll take
`
` 4 breaks periodically, but if at some point you need to
`
` 5 take a break for some reason, please let me know and
`
` 6 I'll do my best to accommodate that as quickly as
`
` 7 possible. I will ask, however, that we not take
`
` 8 breaks while there's a question pending to you.
`
` 9 Do you understand?
`
`10 A. Yes.
`
`11 Q. Okay. Where are you currently employed,
`
`12 Mr. Kumar?
`
`13 A. Google.
`
`14 Q. And where physically do you work at
`
`15 Google?
`
`16 A. New York office.
`
`17 Q. And is that located at -- where in New
`
`18 York City is that located?
`
`19 A. It is Chelsea. 15th Street and Eighth
`
`20 Avenue.
`
`21 Q. And how long have you been employed at
`
`22 Google?
`
`23 A. More than 14 years.
`
`24 Q. So since around 2005?
`
`25 A. Yeah.
`
`7
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 10 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 Q. What is your current job title at
`
` 3 Google?
`
` 4 A. I'm distinguished scientist.
`
` 5 MR. DAVIDOFF: Could you maybe just speak
`
` 6 up a little bit.
`
` 7 Q. When you first started at Google in
`
` 8 2005, what was your job title or function?
`
` 9 A. Research scientist.
`
`10 Q. And did you have a particular area of
`
`11 focus or application?
`
`12 A. Yes. I work in machine learning.
`
`13 Q. And at some point, did your job title
`
`14 change from research scientist?
`
`15 A. I have been promoted, so these are
`
`16 different levels of designations.
`
`17 Q. And has your focus always been in
`
`18 machine learning or have there been various different
`
`19 areas over time?
`
`20 A. It has been primarily, machine learning.
`
`21 But also, allied areas like computer vision.
`
`22 Q. I assume you went to university
`
`23 somewhere.
`
`24 So starting with that, can you tell me
`
`25 where you went to university.
`
`8
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 11 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 A. I did my Ph.D. at Carnegie Melon, New
`
` 3 York City.
`
` 4 Q. And where did you do your undergraduate?
`
` 5 A. I did my undergrad in India.
`
` 6 Q. What was the focus of your Ph.D.?
`
` 7 A. It was machine learning and computer
`
` 8 vision.
`
` 9 Q. When did you receive your Ph.D.?
`
`10 A. 2005.
`
`11 Q. And was Google your first job after your
`
`12 Ph.D.?
`
`13 A. Yes.
`
`14 Q. Have you ever worked anywhere else
`
`15 professionally other than Google?
`
`16 A. I have.
`
`17 Q. Where else?
`
`18 A. Before Ph.D., I was with National
`
`19 Robotics Engineering Consortium. It was affiliated
`
`20 with CMU before starting my Ph.D.
`
`21 And before that, I worked at National
`
`22 University of Singapore in Department of Surgery.
`
`23 Q. What did you do at the National
`
`24 University of Singapore?
`
`25 A. We were developing a robot.
`
`9
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 12 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 Q. At some point during your time at
`
` 3 Google, did you have any involvement with what's
`
` 4 generally referred to as the Content ID system that's
`
` 5 used at YouTube?
`
` 6 A. Yes.
`
` 7 Q. When was the first time you had some
`
` 8 involvement or connection to Content ID?
`
` 9 A. It was a while ago. Somewhere around
`
`10 2014-ish, we started a collaboration on designing a
`
`11 new system.
`
`12 Q. And when you say designing a new system,
`
`13 was it your understanding there was already an
`
`14 existing Content ID system?
`
`15 A. There was an existing match system. I
`
`16 don't know if it was called Content ID system.
`
`17 Q. And you said existing match system; is
`
`18 that right?
`
`19 A. That's right.
`
`20 Q. Okay. Did you have any involvement with
`
`21 the pre-existing match system?
`
`22 A. No.
`
`23 Q. Did you develop any familiarity with the
`
`24 operation of the pre-existing system?
`
`25 A. We had few discussions, but I did not
`
`10
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 13 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 personally dive into the details of the system. I
`
` 3 knew big picture of how the system worked.
`
` 4 Q. Was it your understanding that the
`
` 5 pre-existing system used locality sensitive hashing
`
` 6 to identify matches?
`
` 7 A. Yes.
`
` 8 Q. Did you have a particular point of
`
` 9 contact or someone you interfaced with to gain some
`
`10 understanding of the pre-existing system?
`
`11 A. We did not really go into details of
`
`12 that system. I talked to few engineers, but it was a
`
`13 very quick discussion because they were just unhappy
`
`14 with the system in general.
`
`15 Q. What was the -- can you explain for me
`
`16 the motivation to design a new system.
`
`17 MR. SAFTY: Objection to the extent it
`
`18 calls for speculation.
`
`19 A. They wanted a more accurate system.
`
`20 Q. And what do you mean by more accurate;
`
`21 can you explain what that connotes.
`
`22 A. Which can find the content reuse in a
`
`23 more effective way.
`
`24 Q. And what was your role in the design of
`
`25 a new system?
`
`11
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 14 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 A. I was the technical lead of -- from
`
` 3 machine learning side and research, and I was guiding
`
` 4 the big picture and the experiments, but I did not go
`
` 5 into implementation.
`
` 6 Q. Did that new system have a name or a
`
` 7 project designation of some kind?
`
` 8 A. It was called Siberia.
`
` 9 Q. And you said you started working on that
`
`10 sometime around 2014?
`
`11 A. That's where initial discussions
`
`12 started.
`
`13 Q. Was there some person who was sort of
`
`14 the lead on the Siberia project to develop this new
`
`15 system?
`
`16 A. This project had a lot of people. I
`
`17 talked to many people. Are you looking for specific
`
`18 names?
`
`19 Q. Yes.
`
`20 A. I talked to, for example, Hanna Pasula,
`
`21 Matthias Konrad, and many other engineers there.
`
`22 Q. Did you ever deal with Johan Granstrom?
`
`23 A. Yes.
`
`24 Q. What was his role?
`
`25 A. He was one of the designers of the --
`
`12
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 15 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 one of the engineers working from the Content ID
`
` 3 side.
`
` 4 Q. And what was the sort of approach to
`
` 5 this new system that made it different?
`
` 6 MR. SAFTY: Objection. Vague and
`
` 7 ambiguous.
`
` 8 A. Let me understand what question means.
`
` 9 Q. Sure. So what -- the old system, I
`
`10 think you indicated was, generally speaking, a
`
`11 locality sensitive hashing-based system.
`
`12 What was the new approach that was being
`
`13 implemented in the Siberia system?
`
`14 MR. SAFTY: Same objection.
`
`15 A. Do you want me to go through the
`
`16 technical details?
`
`17 Q. Well, let's start at a high level.
`
`18 Is there a way you would be able to
`
`19 describe what was new or different about it at a
`
`20 general level?
`
`21 A. So it was not based on locality
`
`22 sensitive hashing.
`
`23 Q. The new system?
`
`24 A. The new system. New system was based on
`
`25 primarily, quantization.
`
`13
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 16 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 Q. What do you mean by quantization?
`
` 3 A. You take a value, and instead of keeping
`
` 4 the full position, you quantize it in buckets.
`
` 5 Q. And can you give me an example or
`
` 6 explain how that works in a general sense.
`
` 7 A. Suppose we have a value 4.3, but instead
`
` 8 of keeping 4.3, we'll keep 4, 5, 6, and it will be
`
` 9 mapped to 4. If it is 4.7, let's say it will be
`
`10 mapped to 5. So that is an example of quantization.
`
`11 Q. And how was quantization used to -- in a
`
`12 general sense, in the new match system, the Siberia
`
`13 system?
`
`14 A. So we call the process
`
`
`
`
`
`
`
`
`
`19 Q. Now, you mentioned
`
`20 Was that an approach that was
`
`21 implemented from the beginning of the development of
`
`22 this Siberia system or did that come along the way?
`
`23 A. It was from very beginning.
`
`24 Q. So what was it that was being
`
`
`
` in the Siberia system?
`
`14
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 17 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 MR. SAFTY: Objection. Vague and
`
` 3 ambiguous.
`
` 4 A. The input was the embedding.
`
` 5 Q. And was that embedding a value or a
`
` 6 vector that was extracted from the underlying videos?
`
` 7 A. So when you say videos, is it video
`
` 8 content, audio content, melody content? There are
`
` 9 many.
`
`10 Q. Fair enough. So let's break that down.
`
`11 So you mentioned I think three
`
`12 categories; video, audio and melody content; is that
`
`13 right?
`
`14 A. That's right.
`
`15 Q. Okay. And those are three separate sort
`
`16 of -- I've heard them referred to as channels that
`
`17 are used for matching in YouTube's Content ID system;
`
`18 is that right?
`
`19 A. Yes.
`
`20 Q. So let's start with video.
`
`21 In the case of video, what's an
`
`22 embedding?
`
`23 A. It is a representation of chunks of
`
`24 videos. It could be a shot, it could be a frame, it
`
`25 could be any entity.
`
`15
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 18 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 Q. And when you say a shot, is that a
`
` 3 collection of similar frames?
`
` 4 A. That will be one way of defining shots.
`
` 5 Q. How would you define shots?
`
` 6 A. It is any chunk of video which you think
`
` 7 should be represented together.
`
` 8 Q. And by chunk, do you mean chunk in time
`
` 9 space?
`
`10 A. That's right.
`
`11 Q. So you could represent an individual
`
`12 frame or you might represent some -- more than one
`
`13 frame together if that time space could be
`
`14 represented by a single embedding; is that fair?
`
`15 A. Yes.
`
`16 Q. In the case of audio, what are the
`
`17 embeddings?
`
`18 A. It's again, a bunch of audio frames and
`
`19 where we learn -- or the system learns a
`
`20 transformation which represents this collection of
`
`21 audio frames.
`
`22 Q. And so the embedding is the
`
`23 transformation that represents that group of audio
`
`24 frames?
`
`25 A. That's right.
`
`16
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 19 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 Q. And are those audio frames in the form
`
` 3 of spectrograms?
`
` 4 A. I do not remember exactly how they
`
` 5 extracted the embeddings, because that was another
`
` 6 team. Our work started after giving these
`
` 7 embeddings.
`
` 8 Q. Okay. And then for melody, which I
`
` 9 think was the third category you mentioned, what was
`
`10 the nature of the embedding for melody?
`
`11 A. Nature of embedding? I'm trying to
`
`12 understand.
`
`13 Q. Sure. You've talked about the videos,
`
`14 the embeddings being representations of a frame or
`
`15 group of frames. For audio, I think you similarly
`
`16 said the embedding was a representation of a frame or
`
`17 a group of frame.
`
`18 What's the embedding for melody?
`
`19 A. Same way as audio. Group of frames,
`
`20 audio frames.
`
`21 Q. So the melody embeddings are
`
`22 representations of a group of audio frames also?
`
`23 A. That's right.
`
`24 Q. So I think you started out by saying
`
`25 that
`
`17
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 20 of 201
`
` were these embeddings; is that correct?
`
`
`
`
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2
`
` 3 A. Yes.
`
` 4 Q. So walk me through, if you could, what
`
` 5 is done
`
` 6 A. So embeddings are seen as a vector in a
`
` 7 space. Do I need to define vector?
`
` 8 Q. That's okay. I think I -- I'll go back
`
` 9 if I need to, but go ahead for now.
`
`10 A. And we take the vector, we
`
`
`
`
`
`
`
`
`
`
`
`
`
`17 Q. And what's the process to, as you said,
`
`
`
`19 A. That's just take the vector. There are
`
`20 various versions of this. One can
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`18
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 21 of 201
`
`?
`
`?
`
`?
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 Q. And are there transformations that are
`
` 3 applied to the embeddings that are used in Content ID
`
` 4 before
`
` 5 A. I don't exactly recall. Maybe in -- out
`
` 6 of these three systems, some form may use it. I
`
` 7 don't remember exactly, but this is a variant, and it
`
` 8 may be used in one of the variants.
`
` 9 Q. Okay. Now, am I correct that the -- in
`
`10 the Siberia Content ID system, the embeddings used
`
`11 are
`
`12 A. That's right.
`
`13 Q. How many
`
`
`
`15 A. For which system are we talking?
`
`16 Q. Well, so let's go piece by piece.
`
`17 So for video, how many
`
`
`
`19 A. I have to remember exactly. There are
`
`20 -- every
`
`
`
`
`
`
`
`
`
`25 Q. So just to clarify, a moment ago you
`
`?
`
`
`
`
`
`
`
`
`
`19
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 22 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
`
`
`
`
` 4 A. Yes.
`
` 5 Q. Okay. And so -- so basically, that
`
`
`
`
`
`
`
` 9 A. That's right.
`
`10 Q. And what's the
`
`
`
`12 A. So again, there are several variants. I
`
`13 can give you general, broad picture of how it is
`
`14 done.
`
`15 Q. Sure.
`
`16 A. So for instance, we have a database of
`
`17 let's say, 1 billion embeddings.
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`20
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 23 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2
`
` 3 Q. And the
`
`
`
`
`
` 6 A. That's right.
`
` 7 Q. Now, going back, we talked about the
`
` 8 vectors and
`
` 9 For audio, what's the -- those are also
`
`10
`
`11 A. That's right.
`
`12 Q. For the embeddings. Okay. And what's
`
`13 the
`
`14 A. I don't remember exactly because it is a
`
`15 parameter in the system, how -- right now it is. But
`
`16 I believe it is
`
`17 Q. And is that the same for melody?
`
`18 A. Melody -- I don't recall in melody what
`
`19 exact
`
`20 is this.
`
`21 Q. When you say it's a parameter in the
`
`22 system, how is that parameter set?
`
`23 A. There is a speed and resource,
`
`24 resources, how many machines we use. The speed of
`
`25 the system and the accuracy. So depending on what
`
` we use, but I won't be surprised if it
`
` for video.
`
` as well.
`
`21
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 24 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 operating point we want, we do experiments with
`
` 3 different primary settings and pick what we want for
`
` 4 a particular system.
`
` 5 Q. Is that something that's set as a
`
` 6 parameter but can be changed within the Content ID
`
` 7 system or is that something that was -- that were
`
` 8 optional parameters, but it was sort of hard designed
`
` 9 into the system once you identified the parameter
`
`10 that you wanted?
`
`11 A. I have to understand the question.
`
`12 Q. Sure. Sure. So you talked about the
`
`13 possibility of experimenting with different
`
`14 parameters based on resources and speed and accuracy,
`
`15 things like that.
`
`16 A. Yes.
`
`17 Q. And what I'm trying to understand is is
`
`18 the parameter, for example, how many
`
`19 something that you experimented with, but once you
`
`20 found the value you what, it was set, sort of as a
`
`21 fixed variable in the Content ID system, or is that a
`
`22 parameter that, for example, can be changed in the
`
`23 particular search or particular coding?
`
`24 Do you understand?
`
`25 A. So when we were doing the experiments,
`
` you use
`
`22
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 25 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 early design of the system, we experimented quite a
`
` 3 lot from our side and our -- in the sense from
`
` 4 research side. And we gave them the best parameters
`
` 5 we thought about, but they do lot of experiments
`
` 6 inside their own team. So I wouldn't be surprised
`
` 7 based on different scenarios they changed these
`
` 8 parameters.
`
` 9 Q. So you don't know how that's set in the
`
`10 actual production system?
`
`11 A. We can check the code and get to know
`
`12 what system it is.
`
`13 Q. And where would you look in the code to
`
`14 know how that parameter is set of
`
`15 are used?
`
`16 A. There are clear config files for each
`
`17 match system.
`
`18 Q. Did you say config?
`
`19 A. That's right.
`
`20 Q. Like C-O-N-F-I-G?
`
`21 A. C-O-N-F-I-G. This is a short form of
`
`22 configuration files.
`
`23 Q. And those config files would set these
`
`24 kinds of parameters?
`
`25 A. That's right.
`
`23
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 26 of 201
`
` for the
`
` is indexed.
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 Q. So you mentioned the
`
`
`
` 4 match system.
`
` 5 After those
`
` 6 what happens next, what's the next system?
`
` 7 A. Next step is index of the database.
`
` 8 Once we know these
`
`
`
`10 Q. And how is that index structured?
`
`11 A. Let me understand your meaning of
`
`12 structured.
`
`13 Q. Sure. You mentioned that you have -- I
`
`14 think it's
`
`15 A. That's right.
`
`16 Q. And those are then -- I guess I want to
`
`17 understand what is it that's indexed?
`
`18 A. So I can give you example.
`
`19 Q. Sure, please.
`
`20 A. So we take one vector. This is the
`
`21 embedding vector.
`
`22 Q. Yes.
`
`23 A. From the database, and we take, let's
`
`24 say
`
`25 example, and then we find the
`
` is that right?
`
` for
`
`24
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 27 of 201
`
`
`
` is that right?
`
` in this example.
`
` correct?
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
`
`
`
`
`
`
`
`
` 6 Q. And so that
`
`
`
` 8 A. With the
`
` 9 Q. So a given embedding would have
`
`10 obviously,
`
`11 A. Yes.
`
`12 Q. And would they all have
`
`
`
`
`
`15 A. No. Let me understand your question
`
`16 about location. What do you mean by location here?
`
`17 Q. Well, so you have the
`
`
`
`
`
`20 A. Yes.
`
`21 Q. I guess I'm trying to understand what
`
`22
`
`23 A. These
`
`24 Q. So each chunk has
`
`
`
` Then I assume you have
`
` and so on.
`
` correlates to. Is it --
`
`are independent.
`
`25
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 28 of 201
`
` what is
`
` That's what I'm trying to understand. Is
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
` 2 A. That's right.
`
` 3 Q. Okay. And so are
`
` 4 indexed for? So in other words -- strike that.
`
` 5 Is the index built from
`
`
`
` 7 A. Each
`
`
`
` 9 Q. And what's the -- what's the nature
`
`
`
`11 it --
`
`12 A. These are
`
`13 Q. And what's the relationship between
`
`
`
`15 A. I have to understand this question. You
`
`16 are looking for how do we
`
`
`
`18 Q. Yes. That's what I'm -- I guess what
`
`19 I'm trying to understand is what is it that
`
`
`
`
`
`
`
`23 A. So it
`
`
`
`25 if a
`
`
`
` For example,
`
`26
`
`SANJIV KUMAR-OUTSIDE ATTORNEYS ONLY
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-10 Filed 11/12/20 Page 29 of 201
`
` 1 CONFIDENTIAL - October 15, 2019 - CONFIDENTIAL
`
`
`
`
`
` 4 Q. So the
`