[00:00:00] Speaker 00:
Our third case this morning is 23-21-19, NRA-Google.

[00:00:06] Speaker 00:
OK, Mr. Argyle.

[00:00:08] Speaker 02:
Good morning, Your Honors.

[00:00:09] Speaker 02:
May it please the Court?

[00:00:11] Speaker 02:
Arthur Argyle for Google.

[00:00:13] Speaker 02:
A claim limitation here requires a particular kind of threshold that is determined based on a time series of residuals.

[00:00:22] Speaker 02:
And in this context, a residual refers to a comparison of a measured value to an expected or average value.

[00:00:31] Speaker 02:
The board here found that the prior ARCUDIS reference taught this claimed threshold.

[00:00:36] Speaker 02:
But that was wrong.

[00:00:37] Speaker 02:
CUDIS Nowhere describes determining any threshold based on... Could you speak up just a little bit, please?

[00:00:43] Speaker 02:
Yeah.

[00:00:44] Speaker 02:
Apologies, Your Honor.

[00:00:45] Speaker 02:
CUDIS Nowhere teaches determining any threshold based on the residual values themselves.

[00:00:51] Speaker 02:
Rather, CUDIS taught a static threshold that was just based on a minimum level of frequency that a user selects.

[00:00:59] Speaker 02:
That's at A405 paragraph 96.

[00:01:03] Speaker 02:
Even if that threshold were understood to teach a residual threshold, it would still not teach the claim limitation required here, which requires that the threshold is determined as a function of the residual.

[00:01:16] Speaker 02:
Does it really matter that a residual is based on frequency?

[00:01:22] Speaker 02:
In this case, it's the threshold that's based on the frequency, Your Honor, in CUDIS.

[00:01:27] Speaker 02:
But it does matter that the threshold here is determined based on the residual.

[00:01:32] Speaker 02:
And I think Google's specification is helpful here in illustrating the difference between a static threshold, like what's taught in Cudis, and a threshold determined based on the residual, as in Google's claimed invention.

[00:01:45] Speaker 02:
So at A57, Google's specification compares these two types of thresholds.

[00:01:50] Speaker 02:
And it shows these two types of thresholds at A71 in figure four.

[00:01:56] Speaker 02:
which the board discussed.

[00:01:59] Speaker 02:
In the top graph, you see a static threshold, like what's described in CODIS.

[00:02:04] Speaker 02:
The threshold is just a straight, horizontal line that needs to be exceeded in order for something to happen.

[00:02:13] Speaker 02:
The bottom graph, on the other hand, shows a threshold determined based on the residual, like in Google's claims.

[00:02:20] Speaker 02:
And in the bottom graph, you see that the threshold, shown as a dark, solid line,

[00:02:25] Speaker 02:
varies over time.

[00:02:27] Speaker 02:
It varies because the residual changes.

[00:02:30] Speaker 02:
And here the residual is corresponding to the search activity for something like NYC train outage.

[00:02:39] Speaker 02:
And you see that in the dotted line.

[00:02:41] Speaker 02:
There's a spike corresponding to when a lot of users are searching for NYC train outage, presumably because there's been some issue in the New York City train system that has generated a lot of activity.

[00:02:54] Speaker 02:
So when that spike happens, the residual increases.

[00:02:58] Speaker 02:
It increases because all of a sudden, the measured number of searches is much higher than the predicted level of searches.

[00:03:05] Speaker 02:
And so that leads to a larger residual.

[00:03:07] Speaker 02:
And then importantly here, that leads to a higher threshold.

[00:03:11] Speaker 02:
You see that in the bottom of figure four, where after this spike in activity, the threshold goes up.

[00:03:17] Speaker 02:
Again, because the threshold is determined based on the residuals.

[00:03:22] Speaker 02:
The inventors discovered

[00:03:24] Speaker 02:
that this way of determining the threshold, but do earlier detection in spikes in activity and fewer triggers when the activity or increase is predictable.

[00:03:37] Speaker 02:
Turning back to the Cudis reference, Cudis describes nothing like that.

[00:03:42] Speaker 02:
It just teaches setting a static threshold based on the minimum level of frequency which a user can select.

[00:03:48] Speaker 00:
So the board didn't read it that way.

[00:03:54] Speaker 00:
And looking at CUDAS, it seems to me that interestedness depends on residuals.

[00:04:03] Speaker 02:
We agree that CUDAS does teach a residual calculation.

[00:04:08] Speaker 02:
And that's later in the CUDAS reference at A412.

[00:04:13] Speaker 02:
But importantly, in this discussion, CUDAS doesn't describe any kind of threshold, let alone a threshold required in the claims, which is determined based on the residual.

[00:04:24] Speaker 00:
It says at the end of 96, set a threshold of interestedness.

[00:04:31] Speaker 02:
That's correct, your honor.

[00:04:32] Speaker 02:
And in that paragraph, it goes on to say, the threshold of interestingness, meaning a minimum level of frequency set by the user.

[00:04:40] Speaker 02:
That's not teaching.

[00:04:42] Speaker 00:
Yeah, but elsewhere.

[00:04:44] Speaker 00:
Because I think you just agreed interestedness depends on residuals.

[00:04:50] Speaker 02:
The residual calculation, I agree.

[00:04:53] Speaker 02:
Kudis describes such a residual calculation.

[00:04:55] Speaker 02:
But it never describes how to convert that residual score into a threshold, as in Google's invention.

[00:05:03] Speaker 02:
There's nothing like what is shown in that bottom graph in figure four, where the threshold varies over time because of the change in the residual.

[00:05:14] Speaker 02:
The only threshold taught in Kudis at paragraph 96 is just where a user specifies some minimum level that needs to be achieved.

[00:05:23] Speaker 00:
Well, you're reading that parenthetical as limiting the calculation of interestedness, but elsewhere, the specification makes clear that interestedness can depend on residuals.

[00:05:35] Speaker 00:
That seems to me the heart of the issue, isn't it?

[00:05:38] Speaker 02:
I disagree that's the heart of the issue.

[00:05:41] Speaker 02:
Whether or not interestingness can be evaluated through a residual score is separate from the question of whether CODIS has any teaching of determining the threshold based on the residual.

[00:05:52] Speaker 02:
There's nothing like what Google describes, for example, at 850.

[00:05:55] Speaker 02:
If it weren't for the parenthetical, you wouldn't have an argument, right?

[00:06:01] Speaker 02:
I think we would still have an argument.

[00:06:03] Speaker 02:
Because Kudis, whether you consider just the paragraph 96 disclosure or the separate disclosure of a residual, in neither of those teachings does Kudis describe how to determine the threshold based on the residual.

[00:06:16] Speaker 02:
It's completely unlike what Google's specification sets out, where

[00:06:22] Speaker 02:
At A56, it provides a formula for how to determine the threshold based on the residual.

[00:06:28] Speaker 02:
And that's just nowhere in CODIS.

[00:06:30] Speaker 02:
And because CODIS does not describe anywhere determining the threshold based on the residual, the board's obviousest conclusion lacks substantial evidence.

[00:06:45] Speaker 02:
I think this point is really important here, Your Honor.

[00:06:47] Speaker 02:
And just maybe to put it another way, the threshold in Google's invention

[00:06:52] Speaker 02:
is a function of the residual.

[00:06:54] Speaker 02:
It's not just a constant value.

[00:06:57] Speaker 02:
Whether or not what's set forth in the parenthetical of paragraph 96, Kudis just doesn't describe any way to determine the threshold based on the residual, such that the threshold will vary as the residual varies.

[00:07:16] Speaker 02:
So we think that alone is sufficient to reverse the board's decision.

[00:07:21] Speaker 02:
But unless Your Honors have further questions on that issue, I'd like to now turn to the question of motivation.

[00:07:29] Speaker 00:
So let's go back to my original question.

[00:07:32] Speaker 00:
Why is it that a frequency is different from a residual?

[00:07:37] Speaker 02:
Your Honor, I think this is discussed at A12, where the board sets out a definition for what the residual is.

[00:07:43] Speaker 02:
And the residual requires a comparison of a measured value, such as in Google's specification,

[00:07:50] Speaker 02:
the number of search queries for NYC training.

[00:07:54] Speaker 02:
You compare the measured value of how many searches you're actually getting versus what searches would be predicted through some model.

[00:08:05] Speaker 02:
So it requires a comparison, a difference.

[00:08:09] Speaker 02:
A frequency is just the rate of something over time.

[00:08:11] Speaker 02:
That's just saying, how many searches did we get over this time frame?

[00:08:18] Speaker 02:
So these are just two entirely different

[00:08:20] Speaker 02:
quantities that one does not teach the other.

[00:08:24] Speaker 02:
And I don't think there's really any dispute that a frequency is not a residual.

[00:08:35] Speaker 02:
So I'll turn now to the issue of combining embodiments, which is a separate basis for reversal here.

[00:08:42] Speaker 02:
So as we've been discussing, CODIS has one embodiment with a threshold, but no residual.

[00:08:47] Speaker 02:
That's a paragraph 96.

[00:08:49] Speaker 02:
And then a separate embodiment with residuals, a disclosure of a residual, but no threshold.

[00:08:54] Speaker 02:
That's at paragraphs 185 to 190 at A412.

[00:08:57] Speaker 02:
The board relied on this later embodiment for the teaching of a residual, but it did not make any findings of motivation to combine or reasonable expectation of success.

[00:09:10] Speaker 02:
And that was contrary to this court's precedent.

[00:09:13] Speaker 02:
Because when combining multiple embodiments from a single reference, the board must still find and articulate

[00:09:19] Speaker 02:
a motivation to combine, and a reasonable expectation of success.

[00:09:22] Speaker 02:
And the board just did not do so here.

[00:09:26] Speaker 02:
It took the threshold from the first embodiment and the residual from the second embodiment, but did not make any findings on why the skilled person would be motivated to combine those.

[00:09:39] Speaker 02:
That was particularly improper here.

[00:09:40] Speaker 02:
because these two embodiments are described separately as using conflicting approaches.

[00:09:47] Speaker 02:
So, as discussed with respect to the threshold, the threshold is only described as a subjective threshold in paragraph 96.

[00:09:55] Speaker 02:
The user just selects some level that needs to be achieved.

[00:09:59] Speaker 02:
And again, doesn't determine it based on the value of the residual.

[00:10:05] Speaker 02:
In the later embodiment, CODIS departs from that subjective approach.

[00:10:09] Speaker 02:
It sets forth a statistical approach that doesn't make any use of a threshold and instead just generates a ranked list of keywords.

[00:10:20] Speaker 02:
At paragraph 185, Koudis explains this.

[00:10:23] Speaker 02:
It says that interestingness is a naturally subjective measure, as what is interesting varies according to the group of individuals it is intended for.

[00:10:30] Speaker 00:
I'm a little confused.

[00:10:34] Speaker 00:
On 405, there's a heading Hot Keywords, and that describes some of what we're talking about, including threshold.

[00:10:42] Speaker 00:
And then they come back to Hot Keywords on 412.

[00:10:45] Speaker 00:
Why are those not the same thing?

[00:10:49] Speaker 02:
So agreed, Your Honor, that both of these embodiments are addressing what to do with keywords.

[00:10:56] Speaker 02:
But they differ in the critical respect here, which is that one uses a subjective threshold, whereas one does not use any kind of threshold.

[00:11:04] Speaker 02:
instead uses a more objective statistical approach that just involves creating a list of keywords that users can then include.

[00:11:12] Speaker 02:
And paragraph 186 at 412 sets forth this difference in approach.

[00:11:19] Speaker 02:
It says, given the difficulty and subjective nature of the task, the present invention may adopt a statistical approach to the identification of hot keywords.

[00:11:28] Speaker 02:
And again, in that entire embodiment at 185 to 190, there is no mention of a threshold.

[00:11:35] Speaker 02:
In other words, the solution identified in Kudis here, in this later embodiment, a ranked list of keywords, is directly in response to the disadvantages of what Kudis is describing with respect to the subjective threshold.

[00:11:50] Speaker 02:
So it was improper for the board to just conflate these two embodiments, where the reference itself distinguishes them as using conflicting approaches.

[00:12:00] Speaker 02:
And again, the board made no findings at all on the question of motivation and reasonable expectation of success.

[00:12:06] Speaker 02:
And that was particularly improper here, given the significant differences between these two embodiments, which used incompatible approaches.

[00:12:18] Speaker 02:
Unless your honors have more questions on that, I'll turn now to this last issue I'd like to address.

[00:12:22] Speaker 02:
The board stated that the claims here do not exclude the threshold from also being based on a frequency.

[00:12:32] Speaker 02:
But that finding is not

[00:12:35] Speaker 02:
sufficient to support the board's decision.

[00:12:38] Speaker 02:
Because what ultimately matters here is whether the reference teaches determining the threshold based on the time series of residuals, not on whether or not the claims exclude a frequency.

[00:12:50] Speaker 02:
So because the CUDA's reference just does not teach determining the threshold based on the time series of residuals, it's irrelevant whether the claims exclude a frequency.

[00:13:03] Speaker 02:
Or in this case, they're just silent on frequency.

[00:13:07] Speaker 00:
You're into your bottle time.

[00:13:10] Speaker 02:
I'll reserve the rest of my time for the bottle.

[00:13:11] Speaker 02:
Thank you.

[00:13:12] Speaker 00:
Thank you.

[00:13:13] Speaker 00:
Mr. McBride.

[00:13:23] Speaker 01:
Good morning, Your Honors.

[00:13:24] Speaker 01:
May it please the Court.

[00:13:26] Speaker 01:
I'm Rob McBride on behalf of the USPTO Act and Director.

[00:13:31] Speaker 01:
Judge Dyke, I'd like to start with your point.

[00:13:32] Speaker 01:
The heart of the case here is whether CUDIS is limited to a threshold based on

[00:13:37] Speaker 01:
a frequency, and it's not.

[00:13:40] Speaker 01:
If you read to it, it talks about determining whether the ranked list of keywords are above a set measurement.

[00:13:50] Speaker 01:
And then it goes into one embodiment where it says that threshold can be based on frequency, and that's just one parenthetical.

[00:13:56] Speaker 01:
But then it goes and teaches another embodiment,

[00:13:58] Speaker 01:
where the interesting measure is based on surprise, which is undisputed.

[00:14:03] Speaker 01:
That's based on a series of residuals.

[00:14:07] Speaker 01:
And then it says you can calculate a list of keywords based on that objective measure of surprise.

[00:14:13] Speaker 01:
And then it says you can calculate a list of hot keywords.

[00:14:16] Speaker 01:
which Kudis says is above that set measurement, which in the second embodiment is going to be the threshold based on interestingness based on surprise, which is based on a series of residuals.

[00:14:27] Speaker 01:
And so I think, like you said, Kudis is not limited to that one parenthetical where it gives one embodiment where the threshold is based on frequency.

[00:14:37] Speaker 01:
It also more broadly teaches a threshold based on surprise, which is undisputed, is based on a series of residuals.

[00:14:47] Speaker 01:
I'd also like to point out just the standard of review here.

[00:14:53] Speaker 01:
The dispute over what CUDIS teaches is a question of fact, which is reviewed for substantial evidence.

[00:14:59] Speaker 01:
So even if you were to find that Google's narrow interpretation of CUDIS was plausible, that still wouldn't be sufficient for them to succeed here on appeal.

[00:15:08] Speaker 01:
They also have to prove to you that the board's broader interpretation of CUDIS was implausible or wrong.

[00:15:15] Speaker 01:
And I don't think they can meet that burden here.

[00:15:21] Speaker 01:
If there's no questions on that first argument, I'll touch on the second argument where Google argues that the claims also require this claim construction argument that CUDIS doesn't teach or suggest a system that automatically determines a threshold.

[00:15:37] Speaker 01:
And the director's point of view is that argument is new.

[00:15:40] Speaker 01:
It was raised for the first time on appeal.

[00:15:43] Speaker 01:
So it's forfeited.

[00:15:44] Speaker 00:
And they point to the use of the word manual in, what is it, on 328.

[00:15:54] Speaker 00:
They say they raised it.

[00:15:57] Speaker 01:
They do make that argument.

[00:15:59] Speaker 01:
I would just point you to, if you look at page.

[00:16:02] Speaker 00:
The only threshold in Cudis is the manual threshold.

[00:16:05] Speaker 01:
Correct.

[00:16:06] Speaker 01:
And that's on A328.

[00:16:07] Speaker 01:
And then it's in the argument section.

[00:16:09] Speaker 01:
And if you look at the context of their argument, they're arguing that CUDIS does not teach the claimed, and this is in quotes, the residual triggering threshold.

[00:16:20] Speaker 01:
So this statement is just one statement.

[00:16:23] Speaker 01:
And it's just arguing that CUDIS doesn't teach a residual triggering threshold.

[00:16:27] Speaker 01:
OK, it doesn't say anything about the claims requiring that the system automatically determine the threshold without any manual or human input.

[00:16:37] Speaker 00:
Even in the patent, there would have to be manual input to tell the computer how to set the parameters.

[00:16:46] Speaker 01:
Exactly.

[00:16:47] Speaker 01:
And also, if you look at what they point to as supporting this argument, they point to the claim language.

[00:16:53] Speaker 01:
But the claim language doesn't require that the system

[00:16:57] Speaker 01:
you know, generate the threshold by itself automatically.

[00:17:00] Speaker 01:
It broadly just says it has to determine what it is.

[00:17:04] Speaker 01:
It's broad enough to encompass human input.

[00:17:06] Speaker 01:
And in fact, the claims require receiving a query, which we know Google agrees that that's input from a user.

[00:17:13] Speaker 01:
And then if you look at pages 852 to 853 of the, I'm sorry, 856 to 857,

[00:17:27] Speaker 01:
This is the specification where they cite to the specification disclosing or teaching this requirement that the system has to determine the input.

[00:17:36] Speaker 01:
It doesn't actually teach that.

[00:17:38] Speaker 01:
It actually says that the threshold is equal to, and this is at page 856 in the second full paragraph starting at line six.

[00:17:48] Speaker 01:
It says

[00:17:50] Speaker 01:
In one embodiment, the triggering threshold equals the median plus x. And x is this tuning parameter times IQR, and that's the interquartile range.

[00:18:02] Speaker 01:
And then it goes on to say that x is a tuning parameter.

[00:18:06] Speaker 01:
And then in the next paragraph starting at line 15, it says that the tuning parameter is set.

[00:18:14] Speaker 01:
so that the system triggers based on what it's set to.

[00:18:18] Speaker 01:
And if you go down to, for example, line 24, it says that one can use human raters and historical data to set the tuning parameter.

[00:18:29] Speaker 01:
So it's broadly teaching that one, like a person, is setting this tuning parameter.

[00:18:33] Speaker 01:
So it's encompassing human input there.

[00:18:36] Speaker 01:
And it goes on to say that there's different ways one can set that tuning parameter.

[00:18:40] Speaker 01:
lower for sports queries or higher for other queries.

[00:18:43] Speaker 01:
So even their own specification in the claims, that claim construction argument, it's broad enough to encompass manual input.

[00:18:50] Speaker 01:
So I think that argument is new.

[00:18:52] Speaker 01:
So it's forfeited.

[00:18:53] Speaker 01:
It's also unpersuasive.

[00:18:55] Speaker 01:
And even if you were to find that their claim construction was somehow correct, that narrow interpretation, the claims here, I think, would still have a problem.

[00:19:03] Speaker 01:
Because at that point, they would simply be automating what was previously done manually.

[00:19:10] Speaker 01:
And under this court's case law, just automating, using a computer to automate what was done manually isn't sufficient to render something non-obvious.

[00:19:19] Speaker 01:
If there's no further questions on that issue, I'm happy to yield the rest of my time.

[00:19:25] Speaker 00:
Thank you.

[00:19:27] Speaker 00:
Mr. Argyle, I guess you've got about two minutes.

[00:19:43] Speaker 02:
I'd like to start with where the patent office began, which is about whether this case all turns on whether or not CODIS is limited to the frequency threshold.

[00:19:55] Speaker 02:
And we just disagree that this case is decided if you find that paragraph 96 could also involve a residual threshold.

[00:20:05] Speaker 02:
That is not correct.

[00:20:07] Speaker 02:
Because even if you replaced or supplemented the threshold in paragraph 96,

[00:20:13] Speaker 02:
with some measure of a residual.

[00:20:16] Speaker 02:
A residual threshold itself is not determining the residual threshold based on the residual.

[00:20:23] Speaker 02:
Kudis nowhere describes how the residual score is used to determine the threshold.

[00:20:30] Speaker 02:
It just teaches that the user specifies some level of the threshold.

[00:20:35] Speaker 02:
So we disagree that reading the later disclosure of a residual into paragraph 96 resolves this case.

[00:20:43] Speaker 02:
The critical issue here is that CODIS does not describe determining the threshold based on the residual.

[00:20:51] Speaker 02:
So we think you can resolve that question and reverse the board's decision without even having to get to the system issue.

[00:20:58] Speaker 02:
We disagree with the Patent Office that the system issue was forfeited.

[00:21:02] Speaker 02:
Google argued at A328, as Judge Jake recognized, that we distinguish the prior art as involving a manual threshold as opposed to a threshold that's set by the system, which is required here.

[00:21:16] Speaker 02:
At A13, the board recognized that argument.

[00:21:20] Speaker 02:
So that argument was not forfeited, but in any event, because

[00:21:26] Speaker 02:
Regardless of the system user distinction, CODIS does not describe determining the threshold based on a residual.

[00:21:33] Speaker 00:
The board's decision should be reversed.