Personalized User Model LLP v. Google Inc.

Filing 584

REDACTED VERSION of 578 Declaration,, of Antonio Sistos in Support of Google's Opposition to PUM's Motion to Exclude Portions of Dr. Edward Fox's Non-Infringement Report [Volume II of II] (Exhibits 2-8) by Google Inc.. (Attachments: # 1 Exhibit 2-4, # 2 Exhibit 5-6, # 3 Exhibit 7 part 1, # 4 Exhibit 7 part 2, # 5 Exhibit 7 part 3, # 6 Exhibit 7 part 4, # 7 Exhibit 7 part 5, # 8 Exhibit 7 part 6, # 9 Exhibit 8 part 1, # 10 Exhibit 8 part 2, # 11 Exhibit 8 part 3, # 12 Exhibit 8 part 4)(Moore, David)

1
1

IN THE UNITED STATES DISTRICT COURT

2

FOR THE DISTRICT OF DELAWARE

3
4

PERSONALIZED USER

)

5

MODEL, LLP,

)

6

Plaintiff,

7

vs.

8

)

GOOGLE, INC.,

9

) CA number 09-525 (LPS)
)

Defendant.

10
11

-

)
-

-

-

-

VIDEOTAPED DEPOSITION OF JAIME CARBONELL

12

WASHINGTON, D.C.

13

NOVEMBER 27, 2012

14

The videotaped deposition of JAIME CARBONELL was

15

convened on Tuesday, November 27, 2012,

16

commencing at 10:05, at the law offices of SNR

17

Denton, located at 1301 K Street, Northwest, in

18

Washington, D.C., before Paula G. Satkin,

19

Registered Professional Reporter and Notary

20

Public.

21

-

-

-

-

-

22
23
24
25

Job No. CS1565706

Veritext Corporate Services
800-567-8658

973-410-4040

2
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

4
CONTENTS

APPEARANCES

1
2
3
ON BEHALF OF THE PLAINTIFF:
JENNIFER BENNETT, ATTORNEY AT LAW 4
5
SNR DENTON
6
1530 Page Mill Rd.
7
Suite 200
8
Palo Alto, CA 94304-1125
9
650.798.0300
10
11
ON BEHALF OF THE DEFENDANT:
12
DAVID PERLSON, ATTORNEY AT LAW
13
JOSH SOHN, ATTORNEY AT LAW
14
QUINN EMANUEL
15
50 California Street
16
22nd Floor
17
San Francisco, CA 94111
18
415.875.6600
19
and
20
MARC FRIEDMAN, ATTORNEY AT LAW
21
QUINN EMANUEL
22
51 Madison Avenue
23
22nd Floor
24
New York, NY 10010
25
212.849.7000

JAMIE CARBONELL

EXAMINATION

BY MR. PERLSON
BY MS. BENNETT
BY MR. PERLSON

7
289
294

3

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

ALSO PRESENT:
T.J. O'TOOLE, Videographer

5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

EXHIBITS
CARBONELL EXHIBIT NO:
PAGE NO:
Ex. 1 - A Personal Evolvable Advisor 93
For WWW Knowledge-Based Systems
M. Montebello
Ex. 2 - '040 patent
111
Ex. 3 - '276 patent
125
Ex. 4 - Personal WebWatcher
137
Dunja Mladenic
Ex. 5 - Syskill & Webert: Identifying 186
interesting sites
Ex. 6 - Learning and Revising User 187
Profiles: The Identification of
Interesting Web Sites
Ex. 7 - Learning Probabilistic User 198
Models
Ex. 8 - '032 patent
201
Ex. 9 - Order
224
Ex. 10 - Collecting User Access
254
Patterns for Building User
Profiles and Collaborative
Filtering
Ex. 11 - Report of Michael I.
276
Jordan, Ph.D.

2 (Pages 2 to 5)
Veritext Corporate Services
800-567-8658

973-410-4040

6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

PROCEEDINGS
(10:05 a.m.)
THE VIDEOGRAPHER: On the record
with disk number one of the video deposition of
Dr. Jaime Carbonell taken by the Defendant in
the matter of Personalized User Model LLP versus
Google Inc. and Google Inc. versus Personalized
User Model LLP, both cases being heard before
the United States District Court for the
District of Delaware, Civil Action Number 09-525
LPS.
This deposition is being held at
the law offices of SNR Denton, located at 1301 K
Street, Northwest, in Washington, D.C., on
November 27th, 2012, at approximately 10:05 a.m.
My name is T.J. O'Toole. I am the
certified legal video specialist. The court
reporter is Paula Satkin. We are both here
representing Veritext of New Jersey.
Will counsel please introduce
themselves and indicate which parties they
represent.
MS. BENNETT: Jennifer Bennett
representing Plaintiff Personalized User Model
and the witness, and with me today I have Marc

8

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

Q. And you understand that you are
testifying under oath as if you were testifying
before a jury; correct?
A. Yes.
Q. And I'm going try to be as clear
as I can today and -- but if I ask a question
that you do not understand, please let me know
and I will do my best to make it more clear.
Okay?
A. Okay.
Q. Now, Dr. Carbonell, you -- you
co-authored a book called Machine Learning and
Artificial Intelligence Approach; is that right?
A. There were three of them in that
series.
Q. Okay. And when -- when was the
first one?
A. I believe it was 1983.
Q. When was the second one?
A. 1986.
Q. And how about after that?
A. There was one more. I don't
recall the date. A year or two afterwards.
Q. It was about 1990; does that sound
right?

7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

Friedman.

1
2
Quinn Emanuel representing Defendant Google.
3
MR. SOHN: And Josh Sohn of Quinn
4
Emanuel also representing the Defendant.
5
THE VIDEOGRAPHER: Thank you.
6
Will the court reporter please swear in the
7
witness.
8
Whereupon-9
JAIME CARBONELL
10
a witness, called for examination, having been
11
first duly sworn, was examined and testified as
12
follows:
13
EXAMINATION BY COUNSEL FOR THE DEFENDANT 14
BY MR. PERLSON:
15
Q. Good morning. Could you state and
16
spell your name for the record?
17
A. It's Jamie Carbonell, J-A-I-M-E,
18
C-A-R-B-O-N-E-L-L.
19
Q. And do you go by Dr. Carbonell or
20
Mr. Carbonell?
21
A. Dr. Carbonell.
22
Q. Okay. And you've been deposed
23
before; correct?
24
A. Yes.
25
MR. PERLSON: David Perlson from

9

A. It could be, or it could have been
a little earlier.
Q. You've been publishing in the
machine learning field since then?
A. Yes, I have.
Q. When was the last time?
A. This year.
Q. What -- what did you publish this
year?
A. The latest paper is one at the
Association of Computing Machinery on
learnability of DNF, disjunctive normal form,
expressions.
Q. What's that?
A. Disjunctive normal form is a -learning when you have different expressions of
the target concept. So maybe an example is the
clearest way to explain it.
Q. Sure.
A. If there is bank fraud, there are
different ways of defrauding the bank. For
example, by pretending to be a customer when you
really aren't. By pretending you have a lot
more in an account than you really do and
withdrawing it. Insider transactions that are

3 (Pages 6 to 9)
Veritext Corporate Services
800-567-8658

973-410-4040

22

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

be safe in here.
Q. That's good to know.
A. A Bayesian world, you can use
information like that without data based on
priors which can be updated if you have other
observations. If you observe that a meteor has
struck somewhere else and a second one has
struck, then the probability that a third one
will strike might be higher than it would have
been had there been no other meteor strikes. In
the frequentist case, you're not allowed to use
the equivalent of a prior. You base it only on
the data. And if there is no data, you
basically cannot provide an estimate.
Q. And -- but mathematically, is -is that probability expressed as a number
between 0 and 1?
MS. BENNETT: Objection. Form.
THE WITNESS: In the frequentist
approach, it is.
BY MR. PERLSON:
Q. I'm sorry.
A. I was trying to answer your
earlier question.
Q. The -- okay. Go ahead.

24

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

of 1, that is -- and that does not mean that a
number of 2 is twice as likely to be of interest
to the user as 1, would that be a probability?
MS. BENNETT: Objection. Form.
THE WITNESS: Sorry. I did not
quite grasp the -- the premise of your question.
If you're talking about if you had
a scale that went from 1 to 10, 1 was the lowest
value and 10 was the highest value -BY MR. PERLSON:
Q. Correct.
A. -- 2 then would not represent
twice as likely as 1 because if 1 is the low end
of the scale, 1 means it's not going to happen.
Q. Okay. Okay. Let's say it's
between 0 and 10?
A. Okay.
Q. And if I assign something a number
1 and in order for that range to be a range of
probabilities, wouldn't it be the case that a
number of 2 would have to be twice as likely to
be -- show the interest of a user in a document
in order for it to be a probability?
MS. BENNETT: Objection. Form.
THE WITNESS: Okay. So, first of

23

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

A. It's not just frequentist and
Bayesian. If you go a little broader, there are
things interpreted like fuzzy logic and other
forms of reasoning with degrees of belief, of
belief propagation, that do not require the
values of the interval to be between 0 and 1.
Q. Okay. So fuzzy logic doesn't
require a number between 0 and 1?
A. Some types of fuzzy logic do not
require that; others do.
Q. Okay.
A. Fuzzy logic is a broad term for
introducing numbers into logic -- degree of
belief into logic.
Q. But in order for there to be a
degree of belief, there has to be some sort of
scale of the -- the degree of likelihood of
interest?
A. Yes, sir. That's right.
Q. So if I had a -- if I had a -numbers that went from 1 to 10 and assigning
something a number of 2 was not -- well, in
order for something to be a probability, would
-- let's say I have a -- I can assign numbers 1
through 10, and if I assign something a number

25

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

all, let's call it a likelihood rather than a
probability.
BY MR. PERLSON:
Q. Okay.
A. Technically speaking, it's hard to
think about probability other than 0-to-1
interval. That's how the math works out.
In probability theory, what you're
saying is correct in the sense that it's a
linear scale. If something has twice the
probability value of another event so long as
it's not 0, it means it's twice as likely to
happen.
If you use likelihoods, the
typical interpretation is the same. So if a
0-to-10 scale, an event has a probability or
likelihood of 1 and the second event has a
likelihood of 2, the second event would be twice
as likely to happen as the first.
It is not required that the scale
be linear, but by convention you assume
linearity unless told otherwise. So anybody's
scheme of likelihood is either linear or they
inform you how to calculate it.
Q. How to -- what do you mean, inform

7 (Pages 22 to 25)
Veritext Corporate Services
800-567-8658

973-410-4040

26

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

you how to calculate it?
A. If it's -- if the scale were not
linear, it could be, for example, on a -- based
on a sigmoid function or something else, then
they would have to provide you that sigmoid
function that says if something has a value of 2
and something else has a value of 1, here's how
you estimate how much more likely the value of 2
is over the value of 1. So in the absence of
providing a function, and I used sigmoid
function as an example of one that is sometimes
used, it would be exactly as you say. It would
be linear.
Q. And then let's say that I had a -a -- likelihood numbers of 1 through 4 where 1
was somewhat likely, 2 was very likely, 3 was
extremely likely, and 4 was a certainty of
likelihood. Would that be -- would those
numbers 1 through 4 be probabilities?
MS. BENNETT: Objection. Form.
THE WITNESS: Okay. So, first of
all, you didn't define the bottom end of the
range -- 0 means unlikely or 0 means impossible?
BY MR. PERLSON:
Q. Let's say 0 means highly unlikely?

28

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

it's called an estimation. It's an approximate
calculation of the value.
Now, a model can have multiple
parameters. It can have parameters that
represent whether they like certain terms,
certain concepts, whether they like certain
sources of documents, whether they like certain
topics within the documents, whether they like
to see documents about the same area they've
seen before and so forth.
The collection of all these
parameters together with a mathematical function
that combines them is the model. And estimating
the parameters is finding or estimating a value,
approximating a value, for each one of these
inputs to the model, as it were, one of these
variables in the model. A parameter is like a
variable. It has a value. And you're
estimating the values.
BY MR. PERLSON:
Q. Is the parameter the value or the
-- or the variable?
A. It's used to mean both, and that
is a cause of confusion, I'm afraid. I wish
that my colleagues had been, let's say, more

27

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

A. So where's the point that it
means -- so neither of the two ends are
definitive, so that cannot be converted directly
into a probability.
Q. So you -A. A probability requires both end
points to be nailed down, to be defined. The
impossible versus the certain.
Q. Now, the patent talks about
estimating parameters. Are you familiar with
that?
A. Yes.
Q. And what does it mean to estimate
a parameter?
MS. BENNETT: Objection. Form.
THE WITNESS: It means to compute
the value of that parameter based on the
information available. That computation can be
inexact. It can be an approximation because the
amount of information available is finite. It's
not all possible likes or dislikes by a user.
It's a finite set of those documents they have
already seen. Given that it's based on partial
observations of how a person would react to a
document rather than the totality, that's why

29

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

discriminating in using it to mean only one of
the two. That would have avoided future -future confusion, but a parameter is used to
mean the value and the parameter is also used to
be the variable.
Q. And is that -- is that how it's
used in -- in the patents, too?
MS. BENNETT: Objection. Form.
THE WITNESS: The patent talks
about estimating the parameters. It really
talks about estimating the values of variables.
BY MR. PERLSON:
Q. Sorry. Were you done?
A. Yeah. I'm done.
Q. And the -- in order to -- to
estimate the values of the variables, is that
done by a calculation?
MS. BENNETT: Objection. Form.
THE WITNESS: It is done -everything is done by a calculation. So an
estimation is a calculation based on the
available data.
BY MR. PERLSON:
Q. And that's the -- that's -- the -you mentioned a Dr. Jordan earlier. Is Michael

8 (Pages 26 to 29)
Veritext Corporate Services
800-567-8658

973-410-4040

258
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

THE VIDEOGRAPHER: Counsel passed
me a note asking me how much time we had left,
and I told her that we've been on the record
6 hours and 10 minutes. I have no idea how much
time is left.
MR. PERLSON: Okay.
MS. BENNETT: 50 minutes.
MR. PERLSON: That's all I need to
know.
BY MR. PERLSON:
Q. The -- okay. So let's go to -okay. And so on page 59, you see it says
"representation for web navigation"?
A. Yes.
Q. And then underneath it, it says,
"The probability distribution of the pages to be
accessed is based on collecting the visiting
patterns of many users."
A. Yes.
Q. And what do you understand the
probability distribution of the pages is that's
referred to there?
A. That is the probability of
navigating from one web page to another web page
by following a link between these pages. In

260
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

-- of hyperlinks in order to be in the same
current state.
So he trades off the order of the
model in order to balance accuracy with -- and
generality, and he mostly does an order M equals
2 model as he states on the second -- in the
middle of the second column. He goes through an
example that I have no need to repeat here.
And so this is essentially a
navigation process, and he shows in a finite
state diagram in Figure 1 on the next page -- he
does that illustration so he can calculate the
probability that you will traverse a certain
link, a certain hyperlink, from one page to
another based on what others have traversed
before.
Q. And that probability is used to
determine the variable TIJ; is that right?
A. That probability may be used to
initialize the variable TIJ. He has -- there's
two parts to this -- to this paper, the part
that we are talking about now and the
entropy-based part which is just before it.
In the entropy-based part that he
-- excuse me. E-N-T-R-O-P-Y. In the

259

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

some cases some links are followed by many
users. Other links may be followed by few
users. Some links may be followed by no users.
Q. And how is that information used
in the -- in Wasfi?
A. The main use that he puts to it is
he builds what's called an order M model. So is
there any problem with the recording?
Q. No.
A. We can just continue?
Q. He said that there was 6 hours
10 minutes, and then he passed a note that said
5 minutes. We were just chuckling -- it seemed
inconsistent, but nothing to do -- sorry.
A. So an order -- an order M model
decides how far back in the sequence of
navigation you look to. So an order 1 model
means that you look at the current page and
where else you go next. An order 2 model is
where you came from, the current page. An order
3 model, where you came from before the last one
you came from and so on. The higher the order
of the model, the more information you have, but
then again, the less generalization because you
must have traversed this particular sequence of

261
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

entropy-based part that comes before it, he
defines TIJ in a different way as a negative log
of probability, rather than the probability
itself. That negative log of the probability is
bounded from -- let's see. The probability is
zero -- is bounded from zero to infinity. So
it's not really a probability in -- in that
particular definition of TIJ on page 59,
column 1.
MR. PERLSON: Okay. I think we
need to take a break.
THE VIDEOGRAPHER: This ends disk
number 4 of the Carbonell deposition. The time
is 6:01:58. Off the record.
(A brief recess was taken.)
THE VIDEOGRAPHER: On the record
with disk number five of the testimony of
Dr. Jamie Carbonell in the matter of
Personalized User Model versus Google. The date
is November 27th, 2012. The time is 6:11:21.
BY MR. PERLSON:
Q. So now, we were discussing the
variable TIJ in Wasfi?
A. Yes.
Q. What is that?

66 (Pages 258 to 261)
Veritext Corporate Services
800-567-8658

973-410-4040

262

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

A. TIJ is meant to be measure of
importance or interestingness of the page -- of
the Ith page the Jth user. Ith page to the Jth
user. In fact, I believe Wasfi says so
explicitly -- let me find it. Yes, column 1,
page 59, just above the formula. However, that
statement is not exactly consistent with his -with his formula. This is sometimes called
stochastic entropy rather than the more
traditional or more commonly used Shannon
entropy. Shannon entropy is minus P log P, and
that is bounded on both ends. This is
unbounded, at infinity.
Q. Which is unbounded?
A. TIJ, the H of PR, which is the
same thing.
Q. So something that is unbounded
cannot be a probability; is that right?
A. That's right. It cannot be
normalized into a probability.
Q. The TIJ variable, that indicates
how much weight a new page should get in a
user's profile when that user accesses that
page; right?
A. That's right.

264

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

zero. And so this pathological case will not
occur with nearly as much frequency as it would
occur if it was just an individual user who had
not traversed that link.
Q. Can the weights of a user-specific
learning machine be -- let me start over.
Can the parameters of a
user-specific learning machine be set based on
formulas that take into account the activity of
other users?
MS. BENNETT: Objection. Form.
THE WITNESS: You're talking about
a learning machine for an individual user or a
learning machine for all users?
BY MR. PERLSON:
Q. A learning machine for an
individual user?
A. A learning machine for an
individual user can take account of behaviors of
other users, but it must also take account of
behavior by this specific user so at least some
of the parameters must be estimated from data
specific to this user, not necessarily all of
them.
Q. So a user-specific learning

263

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

Q. And then -A. So a page that would have zero
probability in this case would have an infinite
value.
Q. Okay. But what -- I don't
understand what the -A. The negative -- the logarithm of
zero is minus infinity. And so if you take
minus the minus infinity, it becomes positive
infinity.
Q. You agree that a probability can't
be a negative number?
A. That's right.
Q. So -A. It also cannot be infinite.
Q. Does the fact that the -- the TIJ
is calculated based on a probability
distribution of pages based on collecting the
visiting patterns of many users affect your view
of whether the -- the variable TIJ is a
parameter of a learning machine or user model?
A. No, not really. The -- the fact
that there are many -- information is collected
about many users means that it's less often the
probability of a particular transition will be

265

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

machine must at least have some parameters that
are specific to that user?
A. Yes.
Q. What if the parameters that are
specific to that user -- well, let me give you
-- let me give you an example of something.
If a -- the system creates a
parameter for the user interest in sports and it
determines by the fact that users -- in
reference to all users, that if you've clicked
on sports pages five times, that that indicates
that you should get a weight of .5 for the
variable interest in sports. Would that -would that be a user-specific parameter?
MS. BENNETT: Objection. Form.
THE WITNESS: So how did you
determine that it should have a weight of .5?
BY MR. PERLSON:
Q. Because you look to see it was
assigned based on the activity of all users,
that if in observing all the users, they saw
that if a user clicked on sports pages five
times, that an appropriate weight was .5.
A. Okay. And then this specific user
also clicked on it exactly five times?

67 (Pages 262 to 265)
Veritext Corporate Services
800-567-8658

973-410-4040

298

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

THE WITNESS: No. Estimating the
parameters is -- always means estimating the
values or weights with those parameters. It is
the case, as I mentioned in answers to both of
you, that the field uses the word "parameter"
more loosely sometimes to mean the variables of
the -- and sometimes to mean the values. And
Refuah does that as well because -- and the way
that is consistent with the Court's construction
is consistent with my report and is consistent
with the claim language is "parameters" mean the
values that are being estimated.
BY MR. PERLSON:
Q. Right. And if you look at 1E, it
refers to estimating a probability PUD that an
unseen document D is of interest to the user U.
Then it goes on to say, "wherein the probability
PUD is estimated by applying the identified
properties of the document to the learning
machine having the parameters defined buy the
User Model." Do you see that?
A. Yes.
Q. So that requires that the learning
machine must actually have the values of the
variables that are defined by the user model;

300
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

ACKNOWLEDGMENT OF DEPONENT
I do hereby acknowledge that I have
read and examined the foregoing of the
transcript of my deposition and that:
(Check appropriate box):
( ) the same is a true, correct and
complete transcription of the answers given by
me to the questions therein recorded.
( ) except for the changes noted in
the attached errata sheet, the same is a true,
correct and complete transcription of the
answers given by me to the questions therein
recorded.

_________________ __________________________
DATE
SIGNATURE

299
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

right?
MS. BENNETT: Objection. Form.
THE WITNESS: That's right.
MR. PERLSON: I don't have any
further questions.
MS. BENNETT: Okay. And we
reserve the right to review the transcript and
provide an errata.
THE VIDEOGRAPHER: This ends disk
number 5 and concludes the testimony of
Dr. Jamie Carbonell in the matter of
Personalized User Model versus Google. The date
is November 27th, 2012. The time is 7:08:47.
Off the record.
MR. FRIEDMAN: Ms. Satkin, you did
a stellar job.
(Signature not waived.)
(Whereupon, at 7:08 p.m., the
deposition was concluded.)
- - - - -

301

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22

CERTIFICATE OF NOTARY PUBLIC
I, Paula G. Satkin, the officer before whom
the foregoing proceedings were taken, do hereby
certify that the witness whose testimony appears
in the foregoing proceeding was duly sworn by
me; that the testimony of said witness was taken
by me in stenotype and thereafter reduced to
typewriting under my direction; that said
proceedings is a true record of the testimony
given by said witness; that I am neither counsel
for, related to, nor employed by any of the
parties to the action in which these proceedings
were taken; and, further, that I am not a
relative or employee of any attorney or counsel
employed by the parties hereto, nor financially
or otherwise interested in the outcome of the
action.
My commission expires November 14, 2015.

23
24
25

________________________________
PAULA G. SATKIN
Notary Public in and for the
District of Columbia

76 (Pages 298 to 301)
Veritext Corporate Services
800-567-8658

973-410-4040

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

ERRATA SHEET
VERITEXT CORPORATE SERVICES
800-567-8658
ASSIGNMENT NO. CS1565706
CASE NAME: Personalized User Model v. Google
DATE OF DEPOSITION: 11/27/2012
WITNESS' NAME: Jaime Carbonell
PAGE/LINE(S)/ CHANGE
REASON
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
____/_______/_________________/__________
________________________
Jaime Carbonell

21
22

SUBSCRIBED AND SWORN TO
BEFORE ME THIS______DAY
OF_______________, 2012.

23
24
25

_______________________
NOTARY PUBLIC
MY COMMISSION EXPIRES__________________

Veritext Legal Solutions
25B Vreeland Road - Suite 301
Florham Park, New Jersey 07932
Toll Free: 800-227-8440 Fax: [phone]
_____________, 2012
To: JENNIFER BENNETT, Esq.
Case Name: Personalized User Model v. Google
Veritext Job Number: 1565706
Witness: Jaime Carbonell
Deposition Date: 11/27/2012
Dear Ms. Bennett:
Enclosed please find a deposition transcript. Please
have the witness review the transcript and note any
changes or corrections on the included errata sheet,
indicating the page, line number, change, and the
reason for the change. Have the witness’ signature
at the bottom of the sheet notarized and forward
errata sheet back to us at the address shown above.
If the jurat is not returned within thirty days of
your receipt of this letter, the reading and signing
will be deemed waived.
Sincerely,
Production Department
Encl.
Cc: DAVID PERLSON, Esq.

77 (Pages 302 to 303)
Veritext Corporate Services
800-567-8658

973-410-4040