`
`Exhibit 30
`(Partially Redacted)
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 2 of 276
`
` 1
`
` 2 UNITED STATES DISTRICT COURT
`
`
`
` 3 SOUTHERN DISTRICT OF NEW YORK
`
`
` 4 CIVIL ACTION NO. 1:14-cv-02396-PGG
`
` 5 - - - - - - - - - - - - - - - - - - - - - - - -
`
` 6 NETWORK-1 TECHNOLOGIES, INC.,
`
` 7 Plaintiff,
`
` 8 vs.
`
` 9 GOOGLE, INC., and YouTube,
`
`10 Defendants.
`
`11 - - - - - - - - - - - - - - - - - - - - - - - -
`
`12
`
`13
`
`14 PROSECUTION/ACQUISITION BAR MATERIALS
`
`15 CONFIDENTIAL OUTSIDE COUNSEL ONLY
`
`16 HIGHLY CONFIDENTIAL SOURCE CODE
`
`17 TRANSCRIPT of the stenographic notes of
`
`18 the deposition of HANNA PASULA in the above-entitled
`
`19 matter, as taken by and before LORRAINE B. ABATE, a
`
`20 Certified Shorthand Reporter and Notary Public of the
`
`21 State of New York and Registered Professional
`
`22 Reporter, held at the offices of Williams & Connolly,
`
`23 LLP, 650 Fifth Avenue, New York, New York, on October
`
`24 4, 2019, commencing at 9:15 a.m., pursuant to Notice.
`
`25
`
`1
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 3 of 276
`
` 1
`
` 2 UNITED STATES DISTRICT COURT
`
` 3 SOUTHERN DISTRICT OF NEW YORK
`
` 4 CIVIL ACTION NO. 1:14-cv-02396-PGG
`
` 5 - - - - - - - - - - - - - - - - - - - - - - - -
`
` 6 NETWORK-1 TECHNOLOGIES, INC.,
`
` 7 Plaintiff,
`
` 8 vs.
`
` 9 GOOGLE, INC., and YouTube,
`
`10 Defendants.
`
`11 - - - - - - - - - - - - - - - - - - - - - - - -
`
`12
`
`13
`
`14 PROSECUTION/ACQUISITION BAR MATERIALS
`
`15 CONFIDENTIAL OUTSIDE COUNSEL ONLY
`
`16 HIGHLY CONFIDENTIAL SOURCE CODE
`
`17 TRANSCRIPT of the stenographic notes of
`
`18 the deposition of HANNA PASULA in the above-entitled
`
`19 matter, as taken by and before LORRAINE B. ABATE, a
`
`20 Certified Shorthand Reporter and Notary Public of the
`
`21 State of New York and Registered Professional
`
`22 Reporter, held at the offices of Williams & Connolly,
`
`23 LLP, 650 Fifth Avenue, New York, New York, on October
`
`24 4, 2019, commencing at 9:15 a.m., pursuant to Notice.
`
`25
`
`1
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 4 of 276
`
` 1
`
` 2 A P P E A R A N C E S:
`
` 3 RUSS AUGUST & KABAT, ESQS.
`
` 4 Attorneys for the Plaintiff
`
` 5 12424 Wilshire Boulevard, 12th Floor
`
` 6 Los Angeles, California 90025
`
` 7 BY: BRIAN LEDAHL, ESQ.
`
` 8 MARC FENSTER, ESQ.
`
` 9 (310)826-7474
`
`10 bledahl@raklaw.com
`
`11 mfenster@raklaw.com
`
`12
`
`13 WILLIAMS & CONNOLLY, LLP
`
`14 Attorneys for the Witness
`
`15 725 Twelfth Street NW
`
`16 Washington DC 20005
`
`17 BY: GRAHAM SAFTY, ESQ.
`
`18 KEVIN HARDY, ESQ.
`
`19 (202) 434-5548
`
`20 gsafty@wc.com
`
`21 khardy@wc.com
`
`22 A L S O P R E S E N T:
`
`23 Demarron Berkley, Esq.
`
`24 Google In-House Counsel
`
`25 James Brady, Videographer
`
`2
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 5 of 276
`
` 1
`
` 2 I N D E X
`
` 3
`
` 4
`
` 5 WITNESS EXAMINATION BY PAGE
`
` 6 Hanna Pasula Mr. Ledahl 4, 117
`
` 7
`
` 8
`
` 9
`
`10
`
`11 E X H I B I T S
`
`12 EXHIBIT PAGE
`
`13 1 Plaintiff Network-1 Technologies,
`
`14 Inc.'s Amended Notice of Rule 30(b)(6)
`
`15 Deposition of Defendants 95
`
`16 2 YouTube Content ID Presentation 117
`
`17 3 Siberia Content ID's New Match
`
`18 System Presentation 131
`
`19 4 Eng News Newsletter Printout 144
`
`20 5 Video 7 Launch Description 151
`
`21 6 Self Evaluation Document 163
`
`22 7 Self Evaluation Document 166
`
`23 8 Match System Roadmap Presentation 176
`
`24
`
`25
`
`3
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 6 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 THE VIDEOGRAPHER: Today's date is
`
` 3 October 4th, 2019. The time is 9:18 a.m.
`
` 4 My name is Jim Brady. I'm the
`
` 5 videographer. We're here today at the law firm
`
` 6 of Williams & Connolly, 650 Fifth Avenue, New
`
` 7 York, New York. We're here today in the matter
`
` 8 of Network-1 Technologies versus Google, et al.
`
` 9 Today's witness' name is Hanna Pasula. All
`
`10 attorneys' appearances will appear on the
`
`11 transcript.
`
`12 We ask now that the court reporter
`
`13 please swear in the witness.
`
`14 H A N N A P A S U L A,
`
`15 Having been first duly sworn by a Notary
`
`16 Public of the State of New York, was
`
`17 examined and testified as follows:
`
`18 EXAMINATION BY MR. LEDAHL:
`
`19 Q. Good morning, Ms. Pasula.
`
`20 A. Good morning.
`
`21 Q. Have you ever had your deposition taken
`
`22 before today?
`
`23 A. No.
`
`24 Q. Okay. Then I'm going to go over a
`
`25 couple of things just sort of to help us hopefully
`
`4
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 7 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 make the process a little smoother.
`
` 3 First of all, do you understand the oath
`
` 4 you took a moment ago is the same as if you were
`
` 5 taking if you were testifying live in court in front
`
` 6 of a judge and a jury?
`
` 7 A. Yes.
`
` 8 Q. Okay. And you understand it carries the
`
` 9 same penalty of purgery as if you were testifying in
`
`10 court?
`
`11 A. Yes.
`
`12 Q. As we go along today, you've probably
`
`13 noticed that there is a court reporter who takes down
`
`14 a transcript of everything that we say. To make sure
`
`15 that that's clear and accurate, I'm going to ask that
`
`16 you do your best to wait until I finish my question
`
`17 to start an answer. Okay?
`
`18 A. That sounds good.
`
`19 Q. Likewise, I'll wait so that I don't
`
`20 speak over your answer. Similarly, if you could do
`
`21 your best to say yes or no as opposed to nods of the
`
`22 head or things like um-hum or uh-uh that are very
`
`23 hard to get accurately in the transcript. I may try
`
`24 to remind you if I notice it, but it's not because
`
`25 I'm being rude. I just want to make sure we have a
`
`5
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 8 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 clear record. Okay?
`
` 3 A. I understand completely.
`
` 4 Q. Okay. If at any time you don't
`
` 5 understand my question, please let me know and I'll
`
` 6 do my best to clarify. Otherwise, if you answer,
`
` 7 I'll assume you understood the question. Fair
`
` 8 enough?
`
` 9 A. Yes.
`
`10 Q. Are you on any medication or is there
`
`11 any other health reason that would prevent you from
`
`12 giving your best and most accurate testimony today?
`
`13 A. No.
`
`14 Q. You are employed at Google; is that
`
`15 right?
`
`16 A. That's right.
`
`17 Q. And what is your job title at Google?
`
`18 A. Staff software engineer.
`
`19 Q. Staff software engineer?
`
`20 A. Yes.
`
`21 Q. And how long have you worked at Google?
`
`22 A. 11 and a half years, I believe.
`
`23 Q. And what area or is there a department
`
`24 or part of Google that you work within?
`
`25 A. I currently work at YouTube.
`
`6
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 9 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 Q. And is there any further subpart of
`
` 3 YouTube that you work in?
`
` 4 A. YouTube creator and -- YouTube creator,
`
` 5 YouTube Content ID.
`
` 6 Q. At some point, have you ever been the
`
` 7 technical lead on video matching quality?
`
` 8 A. Yes.
`
` 9 Q. And are you still in that position?
`
`10 A. No. I am now the lead of the match
`
`11 system.
`
`12 Q. And what is the match system?
`
`13 A. It is a system for identifying and
`
`14 understanding content reuse at YouTube.
`
`15 Q. Did you say content reuse?
`
`16 A. Yes.
`
`17 Q. What do you mean by content reuse?
`
`18 A. By content, I mean it is not anything
`
`19 semantic. It refers specifically to video footage or
`
`20 audio recordings, even like audio recordings of
`
`21 single tracks, and also, to melodies. So we try to
`
`22 identify when these things are used in more than one
`
`23 video.
`
`24 Q. And have you worked -- so is that often
`
`25 referred to as the Content ID system at YouTube?
`
`7
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 10 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 A. No. I would say Content ID is a larger
`
` 3 group that the match system is part of.
`
` 4 Q. And what else falls within Content ID
`
` 5 that you're distinguishing that from the match
`
` 6 system?
`
` 7 A. A lot of like communication with
`
` 8 creators or with the partners. A lot of front end
`
` 9 staff that allows them to upload movies, upload their
`
`10 content, and manage it and so on. Let me think.
`
`11 Some elements of YouTube, trust and safety. At least
`
`12 some -- trust and safety is a different team, but
`
`13 part of Content ID is also about identifying content
`
`14 that might be bad in terms of trust and safety.
`
`15 Q. And trust and safety, does that refer to
`
`16 things like pornography or abusive content?
`
`17 A. Yes. Yes, like also, violent
`
`18 extremists.
`
`19 Q. And --
`
`20 A. There is probably other part of Content
`
`21 ID that I cannot quite remember at this moment.
`
`22 Q. Sure.
`
`23 A. There is a lot of infrastructure, like
`
`24 just the piping of passing videos around, passing
`
`25 information around to people storing it.
`
`8
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 11 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 Q. Sure. Let me ask about a specific
`
` 3 aspect.
`
` 4 So within Content ID more generally,
`
` 5 it's my understanding that there is an aspect that is
`
` 6 often referred to as claiming by content partners.
`
` 7 Do you know what that refers to?
`
` 8 A. Yes.
`
` 9 Q. Okay. Is that part of something that's
`
`10 within your area of responsibility or is that outside
`
`11 the match system part that you described?
`
`12 A. It is a sister team.
`
`13 Q. But not part of your team directly?
`
`14 A. No.
`
`15 Q. And who is the lead on the claiming
`
`16 team?
`
`17 A. Krystof Hodor. Krystof, K-R-Y-S-T-O-F.
`
`18 I think Hodor, H-O-D-O-R. Sorry, I always call him
`
`19 Krystof, and that is also his login name. He leads,
`
`20 among other things, the claiming team.
`
`21 Q. Now, within the match system, are there
`
`22 sub components of that that you use as a sort of
`
`23 organizing principle to describe what parts there are
`
`24 to the match system?
`
`25 A. It has changed over time, but I think
`
`9
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 12 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 most of the time we talk separately about the
`
` 3 infrastructure and the quality. Infrastructure is
`
` 4 more like productionalizing, like making sure that
`
` 5 these things run at scale in a reliable manner. And
`
` 6 quality is about designing the algorithems that power
`
` 7 the match system.
`
` 8 Q. And have you worked on the match system
`
` 9 throughout your time at Google?
`
`10 A. No.
`
`11 Q. So what did you work on when you first
`
`12 started, I think you said, about 11 and a half years
`
`13 ago?
`
`14 A. Product search which was at the time,
`
`15 called Froogle.
`
`16 Q. When did you shift to working with the
`
`17 match system?
`
`18 A. I do not exactly remember. I think it
`
`19 might have been 2011.
`
`20 Q. Was the match system already up and
`
`21 running when you started working on it -- when you
`
`22 shifted to working on it?
`
`23 A. Yes. Very much so.
`
`24 Q. Now, how many -- I know there are lots
`
`25 of changes that happened incrementally, but are
`
`10
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 13 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 there, from your perspective, versions, if you will,
`
` 3 of the match system, that you can identify over that
`
` 4 time period since you started working on it?
`
` 5 A. Yes. One big change that occurred over
`
` 6 the last few years is we completely changed the
`
` 7 infrastructure and redesigned the whole system.
`
` 8 Q. When did that take place?
`
` 9 A. I would say starting in something like
`
`10 2014 and finishing this year.
`
`11 Q. And I've seen some references to a shift
`
`12 to a system that's referred to as Siberia.
`
`13 Is that what you're referring to?
`
`14 A. Yes. Sorry. Yes. The new system is
`
`15 called Siberia.
`
`16 Q. How many versions, if you will, of the
`
`17 system were there before Siberia?
`
`18 A. In the same sense, this is the -- just
`
`19 the second version, but many of the components of the
`
`20 system have changed more often.
`
`21 Q. Was there a name that was used to refer
`
`22 to the first version?
`
`23 A. No.
`
`24 Q. Did you work on the first version, the
`
`25 prior, pre-Siberia version, as well?
`
`11
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 14 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 A. Yes. I was an individual contributor
`
` 3 engineer on the video channel.
`
` 4 Q. Now, I think you just said video
`
` 5 channel; is that right?
`
` 6 A. So as I said, we identify content reuse
`
` 7 in video footage, audio recordings and melody. And
`
` 8 we refer to these three different things as channels.
`
` 9 Q. So -- so video, audio and melody are the
`
`10 channels that are -- I've seen that reference in a
`
`11 lot of documents.
`
`12 Is that typically what are referred to
`
`13 as the channels in the Content ID -- or excuse me, in
`
`14 the match system?
`
`15 A. Yes. Although, I'm not sure I should
`
`16 add this, but engineers are much more sloppy about
`
`17 terminology than lawyers. Channel also has other
`
`18 meanings.
`
`19 Q. Okay. I've seen some references in the
`
`20 code to channels.
`
`21 Does that refer to the video, audio and
`
`22 melody or is that something separate?
`
`23 A. You would have to show me the specific
`
`24 piece.
`
`25 Q. Okay. We can look at that later.
`
`12
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 15 of 276
`
` and a brain-based or neural network-based
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 When you came to the match system, what
`
` 3 within the video channel was your particular area of
`
` 4 responsibility?
`
` 5 A. I spent about a year launching a new
`
` 6 version of the video fingerprint, yeah.
`
` 7 Q. And was that -- so my understanding is
`
` 8 there have been iterations of the video fingerprint,
`
` 9 but is it fair to say that broadly speaking, there
`
`10 have been two kind of versions; there was
`
`
`
`12 version?
`
`13 A. There have been two types of
`
`14 fingerprints, but we actually count the version
`
`15 numbers. And we're now up at Video 8.
`
`16 Q. And Video 8, that's a release that
`
`17 relates to how the fingerprints are generated?
`
`18 A. Yes. Each of these numbers refers to a
`
`19 specific path used -- well, to be more crisp, I think
`
`20 it refers to a specific space in which the -- into
`
`21 which the fingerprints -- which the fingerprints form
`
`22 part of.
`
`23 Q. And you said currently the sort of
`
`24 operative release of that is Video 8?
`
`25 A. Yes.
`
`13
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 16 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 Q. Do you recall when you started working
`
` 3 on Content ID and the match system what the video
`
` 4 release was at that time?
`
` 5 A. I believe the one I was working on was
`
` 6 Video 5.
`
` 7 Q. And that was sometime around 2011?
`
` 8 A. Yes.
`
` 9 Q. And at that time, the fingerprints, am I
`
`10 correct, were
`
`11 A.
`
`12 used to index the video prints.
`
`13 Q. How would you characterize what the
`
`14 fingerprints were at the time you were working on
`
`15 Video 5?
`
`16 MR. SAFTY: Objection. Vague and
`
`17 ambiguous.
`
`18 A. Well, they were generated in somewhat
`
`19 hand crafted way using some features and stored as
`
`
`
`
`
`
`
`
`
`
`
`25 over used word for us.
`
` Also, fingerprint is a very
`
`14
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`one step in the algorithm
`
`
`
`
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 17 of 276
`
` So a whole video would have been described by
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 Q. Sure.
`
` 3 A. These -- here, when I say fingerprint, I
`
` 4 mean a description of a whole video. It would have
`
` 5 been made up of sub fingerprints generated, I do not
`
` 6 quite remember, perhaps,
`
`
`
` 8 a sequence of these.
`
` 9 Q. Okay. So let's talk about one of those
`
`10 sub fingerprints, if you will.
`
`11 You said -- I think you indicated you
`
`12 thought they were generated every --
`
`
`
`14 A. Yes.
`
`15 Q. That's also
`
`16 A. Yes, it is.
`
`17 Q. How were those --
`
`18 (Discussion held off the record.)
`
`19 Q. How were -- I want to focus on a single
`
`20 one of those.
`
`21 A. Yes.
`
`22 Q. What you I think called sub
`
`23 fingerprints.
`
`24 A. Yes.
`
`25 Q. How were each of those generated in that
`
`15
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 18 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 earlier system? Let me step back for a second.
`
` 3 I think you mentioned there were sort of
`
` 4 two types of fingerprints, if you will, that have
`
` 5 been used even though there were iterations of each
`
` 6 of those over time; is that fair?
`
` 7 A. Yes.
`
` 8 Q. Okay. So I want to start with the first
`
` 9 type which was in the earlier time, as I understand
`
`10 it, and was in use, generally speaking, when you
`
`11 started working on the match system.
`
`12 So in that first iteration, how was each
`
`13 of those sub fingerprints, those every four times a
`
`14 second, how was one generated?
`
`15 A. I am actually not completely sure how
`
`16 they were generated for the earlier iterations before
`
`17 I started.
`
`18 For Video 5, it is quite a long process,
`
`19 actually describing. How would you like me to
`
`20 approach it?
`
`21 Q. So I would like to understand what were
`
`22 the -- at a sort of basic level, what is that sub
`
`23 fingerprint?
`
`24 So is each one -- first of all, is each
`
`25 one
`
`16
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 19 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 A. Yes.
`
` 3 Q. Okay. And how was that
`
`
`
` 5 A. I may not be recollecting some things
`
` 6 completely correctly, but from as far as I remember,
`
` 7 it goes like this; first, we would take several
`
` 8 frames, so we're talking about a specific point in
`
` 9 time. We will be generating a sub fingerprint to
`
`10 describe a specific point in time in the video.
`
`11 So first, we would take several frames
`
`12 around that point in time and
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`17
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 20 of 276
`
`
`
`
`
`
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`23 Q. And then all of those things together
`
`24 were essentially calculated out to be represented by
`
`25 that
`
`
`
`
`
`
`
`
`
`
`
`18
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 21 of 276
`
` and the
`
`
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 A. Eventually. Like you want me to go into
`
` 3 detail?
`
` 4 Q. That's okay. I think I have a rough
`
` 5 understanding.
`
` 6 And so each of those that you mentioned,
`
` 7 those were sort of values that were set up for every
`
` 8
`
` 9 collection of them would be what was often referred
`
`10 to as a fingerprint for that particular video?
`
`11 A. We have actually missed several steps.
`
`12 Q. Oh, sorry. Go ahead, please.
`
`13 A. So after collecting these, we would also
`
`
`
`
`
`
`
`17 Q. Collect?
`
`18 A. These descriptors that I have been
`
`19 talking about. So you would have descriptors at this
`
`20 current point in time and then you would have them
`
`21 also from the -- a few times before and after. I do
`
`22 not remember exactly how long that interval was,
`
`
`
`
`
`
`
`
`
`
`
`19
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 22 of 276
`
`
`
`
`
`
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`10 Q. And I just want to make sure I
`
`11 understand.
`
`12 There was -- so there was
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`
`21 A. Would be what I call a sub fingerprint.
`
`22 Q. Okay. So each hash representation is a
`
`23 sub fingerprint in the way you've been describing it?
`
`24 A. The sub fingerprints were generated
`
`
`
`
`
`
`
`
`
`20
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 23 of 276
`
` for a
`
` apart for video,
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 called actually,
`
` 3 Q. Thank you.
`
` 4 What I would like to try to understand
`
` 5 is if I'm thinking about a video as a sequence in
`
` 6 time, how -- is it true, is it the case that there
`
` 7 would be
`
` 8 single video; each one is a sub fingerprint. Am I
`
` 9 understanding correctly?
`
`10 A. For each of these instances in time
`
`11 which I think was
`
`12 there would be one of these
`
`13 representing what was going on around that point in
`
`14 time in a highly encrypted and probabalistic way.
`
`15 Q. Understood. Okay.
`
`16 And the collection of all of those would
`
`17 be what was referred to as the fingerprint for that
`
`18 video?
`
`19 A. Yes. Although, we no longer use that
`
`20 terminology because it's confusing.
`
`21 Q. Fair enough.
`
`22 Now, at some point, am I correct that
`
`23 YouTube switched to a completely different approach
`
`24 to generating fingerprints; if that's a fair way to
`
`25 put it?
`
`21
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 24 of 276
`
` that you were talking about earlier.
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 A. Yes.
`
` 3 Q. And how would you characterize the later
`
` 4 or the second approach?
`
` 5 MR. SAFTY: Objection. Vague and
`
` 6 ambiguous.
`
` 7 A. We switched to a neural network-based
`
` 8 approach.
`
` 9 Q. And which -- so you mentioned that for
`
`10 example, Video 5 I think was still using
`
`
`
`12 Which release was the first to utilize
`
`13 the neural network-based fingerprint?
`
`14 A. I believe it was Video 6.
`
`15 Q. Okay. The neural network-based
`
`16 fingerprinting also -- well, I guess not also. It
`
`17 generated
`
`18 that correct?
`
`19 A. No.
`
`20 Q. Thank you. Sorry.
`
`21 A. And we stopped referring to them as sub
`
`22 fingerprints.
`
`23 Q. What is the reference to them at this
`
`24 point?
`
`25 A. Embeddings.
`
` as the sub fingerprint; is
`
`22
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 25 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 Q. E-M-B-E-D--
`
` 3 A. Yeah. Things embedded in a space, yeah.
`
` 4 It's -- yeah. It's what they would be called in the
`
` 5 literature of neural networks.
`
` 6 Q. And can you describe what each embedding
`
` 7 is.
`
` 8 MR. SAFTY: Objection. Vague and
`
` 9 ambiguous.
`
`10 A. It's a point in a
`
`11 Q. And that's represented by the
`
`
`
`13 A. Yes. It's a vector which is a
`
`14 description of a point in space.
`
`15 Q. And what does that point in
`
`
`
`17 underlying video?
`
`18 A. Not much. It's not so much that it
`
`19 represents something intrinsically. It has sort of a
`
`20 certain property.
`
`21 Q. Well, so when you say it has a certain
`
`22 property, what do you mean by that?
`
`23 A. So we trained the neural network so that
`
`24 embeddings corresponding to -- in video, video frames
`
`25 that came from the same footage, from the same point
`
`represent with respect to the
`
`23
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 26 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 in the same footage, end up close together in the
`
` 3 same space, even if they have been heavily
`
` 4 transformed; and that embeddings that come from
`
` 5 completely different frames, are far away.
`
` 6 Q. So an individual embedding, is that
`
` 7 generated from a snapshot of a point in time within
`
` 8 the video?
`
` 9 A. For the video channel, it's a single
`
`10 frame.
`
`11 Q. And that single frame, how often are
`
`12 those captured to generate the embeddings?
`
`13 A. I believe it is still
`
`
`
`15 Q. So let's step to just one of those.
`
`16 So I take it, then,
`
`17 there is something that essentially extracts a single
`
`18 frame at that point in time from the video, and then
`
`19 that is processed in some way to
`
`
`
`21 A. I believe we just request a transcode
`
`22 that has
`
`23 infrastructure or we ask it to transcode it in such a
`
`24 way and then we process each frame individually.
`
`25 Q. When you say you ask it to process a
`
` from YouTube's
`
`24
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 27 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 transcode, can you just explain to me what that
`
` 3 means.
`
` 4 A. So there are many different video
`
` 5 represent -- a video can be represented many
`
` 6 different ways, with like higher, smaller resolution.
`
` 7 There can be different frame rates, different N
`
` 8 codings of the video. We ask for a specific N coding
`
` 9 which is -- in which the frames are explicit --
`
`10 represented explicitly. And they
`
`
`
`12 YouTube has a lot of infrastructure for
`
`13 purchasing different transcodes for different
`
`14 purposes.
`
`15 Q. And so -- so you receive that transcode,
`
`16 and then it -- that's essentially a collection of
`
`17 frames from the video; is that fair?
`
`18 A. Yes.
`
`19 Q. Okay. And each of those frames is then
`
`20 processed to generate one of these vectors or
`
`21 embeddings?
`
`22 A. For the video channel, yes.
`
`23 Q. And how is it different from the audio
`
`24 channel?
`
`25 A. The audio channel is -- well, based on
`
`25
`
`HANNA PASULA-HIGHLY CONFIDENTIAL
`
`
`
`Case 1:14-cv-02396-PGG-SN Document 241-5 Filed 11/12/20 Page 28 of 276
`
` 1 CONFIDENTIAL - Pasula - October 4, 2019 - CONFIDENTIAL
`
` 2 audio, and audio is a wave form. So what we do is we
`
` 3 produce what is called a spectrogram. It is a visual
`
` 4 representation of the frequencies and the times in
`
` 5 the audio wave form, and it basically sort of shows
`
` 6 the notes that are occurring at different points in
`
` 7 time and how they're sustained. And so we produces
`
` 8 this rolling tape of what's going on with audio, and
`
` 9 then we take snapshots of it and produce embeddings
`
`10 from those.
`
`11 Q. And are those snapshots also
`
`
`
`13 A. I believe they are
`
`14 Q. For audio,
`
`15 A. Yes, and they are
`
`16 Q. How -- what's the duration on the ones
`
`17 for video? Is it --
`
`18 A. It's just a frame.
`
`19 Q. Just the frame.
`
`20 And is there a third way that -- I
`
`21 believe you also mentioned there is a melody channel.
`
`22 A. I am less familiar with how the melody
`
`23 channel preprocess