[00:00:00] Speaker 00:
Next case for argument is 24-2208, Apple v. Zentian.

[00:00:07] Speaker 02:
Good morning. Good morning. Thank you, Your Honor. May it please the Court, Joseph Palmore here on behalf of Apple. The Board made three legal errors, two on obviousness and one on procedure, and they require reversal or at least remand for application of the proper legal standards. First, the board failed to recognize that a single reference and a combination can disclose a claim element. The board recognized at page 21 of its order that Jang alone disclosed the limitations of 1D and 1E.

[00:00:37] Speaker 02:
That should have been enough to move on to motivation.

[00:00:41] Speaker 03:
I guess I come into the argument understanding what the board did as follows.

[00:00:49] Speaker 03:
that what was not sufficiently shown in the petition was that, one way to put it, the combination itself performs all of the elements, so that it's not enough if Zhang alone would teach something. The question is whether Zhang, when put into the combination, would still mean that the combination teaches, this is what, 1D and 1E?

[00:01:24] Speaker 02:
That's right, Your Honor.

[00:01:25] Speaker 03:
And that that was the problem it found, and it didn't seem to me like a methodological error. So what would be left is a factual question. Tell me why I'm wrong in thinking about it that way.

[00:01:41] Speaker 02:
I think the board...

[00:01:44] Speaker 02:
Although it denied it was doing bodily incorporation, that is in essence what it was doing. Because then the key sentence is at page 21 where it says, the petition explains how Jang's pruning algorithm generates an initial score and uses the initial score to prune or, quote, determine whether to continue processing as recited in claim one. So that box is checked. Those claim elements were disclosed. But then the board goes on, and this is where we think it entered into error. but does not explain how Jang's pruning algorithm would be implemented on Chen's four clusters.

[00:02:20] Speaker 02:
That's a question of implementation that really sounds in physical, in bodily incorporation. They wanted to know how would you combine these? How would this work?

[00:02:31] Speaker 03:
I guess I'm thinking back to, I think there are a number of these cases, but the one I associated with is one of the personal web cases. where I think even after a second try, we said in agreement with the board that the petitioner, which might have been Apple, I'm not sure, did not show how a particular claim element was met in the combination, and therefore both for reasonable expectation of success and motivation, The case for skilled artists in getting there was just not made.

[00:03:11] Speaker 03:
So the how is actually relevant, I think, even without doing a bodily incorporation. It's a failure of explanation if you don't illustrate it well enough.

[00:03:24] Speaker 02:
Well, I think in personal web, this court did reverse the board. But that was because the board relied on the wrong reference. There was no suggestion there that a single reference disclosure is insufficient to establish that the combination also discloses. It was just that the board relied on a reference that Apple hadn't even pointed to. And what this court said was, The board – Apple relied solely on – I think it was the Aramelli reference. It was a single reference there and sent it back to redo looking at the reference that Apple had actually relied on in the record.

[00:04:00] Speaker 02:
There was no suggestion that reliance on that single reference was inadequate. I think Your Honor is right that, of course, there is kind of a how and why question, but that's on the motivation to combine references.

[00:04:11] Speaker 02:
side of the analysis, this was the first part of the board's decision where it was really doing the mapping. And it was looking at each element and saying, where is this element disclosed in the prior art? And for these, 1D and 1E, it was undisputed that the Jang tree algorithm disclosed these functional elements. That should have been enough to kind of check that box and move on.

[00:04:37] Speaker 03:
It keeps seeming to me that there's a...

[00:04:40] Speaker 03:
a little leap that you're taking from mapping to particular references and saying, now we're done with the disclosure part. I don't quite think you're done with the disclosure part. The question is whether the combination teaches all of the elements. And if it, in particular here, where it is the same, was it processor? Is that what the basic noun is here?

[00:05:08] Speaker 02:
There's a circuit and then there are processors. Right.

[00:05:10] Speaker 03:
has to do all these things. And you didn't show how, when these were combined, all those things were taught.

[00:05:19] Speaker 02:
Right. I want to come back. I think we did actually show that, and I want to come back to that. I understand your question, but I read Intel versus PACT, and I apologize. There are two Intel versus PACTs that are in this case, both from 2023. This is the unpublished Intel versus PACT. It has this same kind of structure where there was basically a single reference disclosed one claim element and a different reference disclosed others, and the board said, like here, You didn't show how they would have been combined.

[00:05:50] Speaker 02:
And what this court basically said was, if a single reference discloses it, then the combination discloses it.

[00:05:57] Speaker 03:
And they sent- You said that's non-precedential.

[00:06:00] Speaker 02:
It is non-precedential, but I think it applies. And it applies Enray-Etter, which, of course, is one of the foundational cases in this area, which says that bodily incorporation or implementation details between references for a combination is basically irrelevant. in establishing obviousness. But I do want to go to your, this is the second argument, Your Honor, which is that Apple actually did disclose how the combination would function. And the board erred by saying that this was a new theory that was advanced for the first time on reply.

[00:06:35] Speaker 02:
Apple was quoting verbatim from its petition when it made this argument. So a It was not, even remotely, a new theory.

[00:06:45] Speaker 02:
And this is at page 90 of the appendix, which is the Apple petition. And this was, again, quoted in the reply.

[00:07:02] Speaker 02:
Apple said at the bottom last paragraph, each of the one or more clusters in the modified Chang-Chen circuit would have been configured by Aposita to include a plurality of processors described by Chen performing the techniques described regarding the TreeSearch Engine 74. TreeSearch Engine 74 is used throughout the petition as what maps onto 1D and 1E, this algorithm. So Apple did explain that the processors in the clusters would do this work.

[00:07:32] Speaker 02:
And so the idea, so we don't think that implementation details were required, but if they were, they were there. And this was clearly not a new theory because if you look at the relevant paragraph from the reply, and we reproduce it in our blue brief, it's all quotes and citations to the petition. There's nothing new in there.

[00:07:53] Speaker 03:
So this was not... Am I remembering right that... There's an issue about whether there were citations to the petition's discussion of 1D and 1E as opposed to other parts of the petition.

[00:08:05] Speaker 02:
So what the board said was, well, those were the 1A, 1B, and 1C. Those were quotes from 1A, 1B, and 1C. But just to take a step back, what the board's criticism was, and under this patent and this claim, there's kind of a processor architecture part. That's 1A, 1B, and associated functional limitation in 1C. And then there are D and E, which are purely functional. It doesn't even say they're done by the processors. They're just done by the circuit. right? So what the board said was you didn't show how the D and E functions would have been implemented in the processor architecture of 1A and 1B.

[00:08:43] Speaker 02:
So Yes, this explanation that I pointed you to is in the 1A part, but it stands to reason that if the question is, okay, how would this architecture have performed these functions, I guess you could have put it in one of two places, either in the 1A architecture discussion or in the 1D through 1E functional section.

[00:09:05] Speaker 02:
Apple put it in 1A.

[00:09:07] Speaker 00:
But why wouldn't you need to at least have a cross-reference in the 1D, 1E to say, go look at 1A?

[00:09:16] Speaker 02:
Well, because I...

[00:09:18] Speaker 02:
Again, going back to kind of the legal point, we just had to show, and it was undisputed, that Jang did disclose and did teach 1D and 1E. And, you know, I don't know that there needed to be a C-Supra site when, if we needed to do more, if the question was how would this processor architecture have implemented the functions, if we had already addressed that very question in the discussion of the processor implementation, that we had already shown it.

[00:09:49] Speaker 02:
If I could go to the final part of this, the motivation to combine discussion, because the board did make a separate finding on that. And here this touches a little bit on some of the colloquy in the earlier case. There was really an axonics problem here because the board focused its motivation to combine analysis and made dispositive a purported failure to – to explain motivation to combine as to an unclaimed element. So on 31 and then in subsequent pages, the board repeatedly criticizes Apple and says, you haven't shown how there could be inner cluster memory access in this combination.

[00:10:30] Speaker 02:
How would the clusters talk to each other for purposes of calculating a pruning threshold is what they were getting at there. That is not claimed, right? Again, if you look at claim one, it claims a speech recognition circuit. It has processors. And we know the processor, the only thing that the job given specifically to the processors is to calculate a probability. That's 1C. Then you look down to 1D and 1E, which are the purely functional algorithmic elements.

[00:11:00] Speaker 02:
Those are not done by the processor. They're done just by the circuit. And this is a comprising claim. So there could be unstated elements here. And so what part of the circuit is actually doing the comparison that might require communication across the memories or across the processors is not recited in claim one. And under axonics and cases like it, if the purported failure to show a motivation to combine goes to an unclaimed element, then it's not a failure of motivation to combine because the invention is, of course, defined by the meets and bounds of the claim.

[00:11:39] Speaker 02:
And the question is, would this have been obvious to satisfy the claim limitations? And the claim limitations don't say anything about the There's the Viterbi algorithm that's discussed, you know, inner cluster memory, and maybe there'd be latency problems with inner cluster memory. And the final point I would make on motivation to combine, and again, this harkens back to the first case, is that this was really, again, a straight up KSR. Intel versus PACT, the other Intel, the published Intel versus PACT case, where it was known in the art, including with respect to speech recognition, that using clustered architecture was a way to improve performance and scalability.

[00:12:23] Speaker 02:
This was known. Our expert talked about this. This is 788 through 790 when he talked about prior art references that were specifically about speech recognition. And then Chen itself discussed Jiang, which was the principal reference, said that it could be implemented with respect to multiple processors. And then Chen said systems using multiple processors could be retrofitted to use this clustered architecture. So this was a known problem to a known solution, and you had two references attacking it in a very similar way, and it would have been obvious to a person of ordinary skill in the art to be motivated to combine them.

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

[00:13:16] Speaker 01:
Thank you, Your Honors. Judge Toronto correctly summarized the problems that the Board identified. And what I think bears significant emphasis in this appeal is the standard of review that is applicable to the issues that Apple contests, which Apple has not even attempted to meet. So the Board has the prerogative to decide what the petition says, what theories are presented in the petition, and what is and is not in the petition.

[00:13:47] Speaker 01:
And then the Board's findings in that respect are subject to abuse of discretion review. And for example, this Court's precedence in Corphtonics v. Apple and Henny Penny v. Frymaster, as well as the Netflix decision and even more recent decisions, all stand for that proposition. So in order to prevail on the question of of whether the board erred on the 1D and 1E findings and whether the reply theory was already included in the petition, Apple would need to show abuse of discretion.

[00:14:20] Speaker 01:
Apple did not even argue abuse of discretion on those issues in its briefs. It did not argue that today. And so it simply falls short of meeting the standard. Substantively, though, the petition speaks for itself. The 1A through C arguments go to a different aspect of speech recognition as recited in the claims. So 1A through C have to do with how the tree search is performed, where an evaluation is made of the so-called phonemes and effectively what is it that the speech could be, what words could it be, and then 1D and 1E go to the question of pruning, which is, okay, now that we've done some of the analysis, we want to go back through the tree and cut out the branches that seem so low probability that they're not worth considering anymore.

[00:15:12] Speaker 01:
So presenting the theory as to 1A through C does not automatically port over by any means to presenting the necessary showing on D and E, and that is substantively what the board found. The board expressly found that.

[00:15:26] Speaker 03:
The board specifically said... How does that distinction between A through C on the one hand and D and E on the other, the... the assignment of probabilities on the one hand and the pruning on the other, either equate to or differ from what the board said about Zhang by itself versus Zhang in the combination.

[00:15:52] Speaker 03:
So how do those two things relate to each other?

[00:15:56] Speaker 01:
They relate to each other in that Zhang teaches a specific pruning algorithm or approach to pruning that is part of what the petition would have to rely on for 1D and 1E. And then as you pointed out, Your Honor, that pruning algorithm would have to be implemented within the speech recognition circuit of the claims, and that's expressly recited in 1D. It has an antecedent reference, the speech recognition circuit. And that speech recognition circuit is, by Apple's own admission, not JANG, because JANG doesn't have a cluster of processors, and the speech recognition circuit of the claims must have a cluster of processors.

[00:16:37] Speaker 01:
And therefore, the pruning algorithm of JANG must be into the clustered processor architecture of Chen under the combination theory of the petition. And this is where the petition was entirely silent and therefore failed to meet its burden.

[00:16:54] Speaker 03:
So I'm not sure I understood what you're saying.

[00:17:00] Speaker 03:
well enough what Mr. Pellmore was saying, but maybe one version of it was this. When you look at the claim, you have a circuit and A, B, and C have processors to do something, and then D and E don't have to use those same processors. There can be other processors.

[00:17:18] Speaker 03:
Rather, the circuit, in some way or other, not specified, has to perform this pruning function.

[00:17:27] Speaker 03:
What's missing, if that's right, than when the petition says, Jang provides for pruning. Stick it in the circuit. It doesn't have to be the same processors. Just stick it in maybe some more processors.

[00:17:43] Speaker 01:
Well, first of all, the petition never even says stick it in the circuit. And second of all, the petition says nothing as to how you would, quote, stick it in the circuit to end up implementing it. It's not a technical term. I understand. But it is sort of the conceptual issue, right, is that you have an algorithm in Jeng and you've got hardware in Chen. And now you have to bring these two together. And whether you're going to run that algorithm on every single processor in every cluster, which is what Apple explicitly said in its reply but never said in its petition, or whether you're going to run that algorithm on only one processor in one cluster, which is what their expert testified at his deposition when he was confronted with the fact that he had no theory on this issue at all, those things have to be explained and then they have to be stress tested.

[00:18:30] Speaker 01:
And the evidence, first of all, neither of those theories nor any other theory as to how you would make the combination was presented as to 1D and 1E. And second of all, the theories that were raised both in the deposition testimony of Apple's expert and in the reply brief for the first time were rejected substantively by the board because they simply fail in view of the evidence to actually arrive at a working combination. So, yes.

[00:18:59] Speaker 03:
You use the expression stress tested. What do you mean? You have to run it and see a good result and say, yay, good result?

[00:19:07] Speaker 01:
No, no, no, no. The stress test of actual litigation, where one side presents its theories and the other side gets to refute them based on evidence. Right. So we didn't even get to the starting point with Apple's theories based on the petition. But when Apple tried to fill that gap, either through the deposition testimony of its expert or through its reply brief theory, we then applied the stress test of litigation of what trial is all about. And we put forth an avalanche of evidence that showed that those theories don't hold up.

[00:19:37] Speaker 01:
And the board's findings are subject to substantial deference review. Abusive discretion review and substantial evidence review on both of those findings, as well as the initial finding that the petition was silent on how to achieve 1D and 1E in the way that the combination of the petition itself required.

[00:19:57] Speaker 01:
Your Honors, I'd like to address some of the points that were made in the opening argument.

[00:20:13] Speaker 01:
I do want to note that 1D and 1E are not purely functional, right? Because they recite the speech recognition circuit. And so this goes back to the question of, is it enough to simply show that Jang teaches an algorithm for pruning? And again, I think Judge Taranto correctly summarized the board's findings, which again are reviewed with deference, that Jang's algorithmic teachings both under the theory of the petition itself and in view of the claims, are inadequate to meet 1D and 1E.

[00:20:51] Speaker 01:
Obviously, the board did make extensive additional and separate findings on the failure of the motivation to combine, and Apple would need to prevail in overcoming the deferential review afforded to all of the board's findings in this case to disturb the board's conclusion.

[00:21:09] Speaker 01:
With respect to the motivation to combine, the board found that Apple rested on unproven assumptions that Zentian effectively disproved, and the board provided a very extensive discussion in detail from Appendix 27 to Appendix 38. I think one very important aspect of the board's finding is set out right at the beginning at Appendix 31 to 32, where the board found that the combination would require processors in each cluster to access memories of all other clusters.

[00:21:42] Speaker 01:
And Apple's counsel today said, well, that's not required by the claims, so we should just ignore that issue. And just to be clear, the problem there is that Chen expressly teaches that none of the processors in any cluster can access all of the memories of the other clusters. So the combination would, in practice and in the way that it would operate, require something that Chen says it can't really do. And even Apple's petition and its expert agreed on that point that Chen teaches it does not have that capability.

[00:22:14] Speaker 01:
This is not immaterial simply because it's not a recited requirement of the claims. It's highly material under this court's case law because the petition or any obviousness theory has to explain how the combination would operate. to meet the claim limitations. And then the motivation to combine is assessed based on that combination that would be required, right? So the question is not simply whether a person of skill in the art would have been motivated to combine some pruning algorithm with some set of processors and memories.

[00:22:49] Speaker 01:
The question is whether they would have been motivated to do what the specific combination detailed. And in this context, reviewing that fact-specific question, the board found that the consequence of the combination would be this need for the processors to access all the memories, which Chen could not do.

[00:23:07] Speaker 01:
and it found a lack of motivation on that basis, as well as its findings that the combination would not have been faster, it would have actually been slower, that there were significant architectural differences between Jung and Chen that would weigh against the combination and cause latencies. There was no evidence of a cost improvement. The combination did not rely on a known technique. There was no scalability benefit, and in any case, Any scalability and flexibility considerations were outweighed by the numerous showings of downsides to the combination.

[00:23:46] Speaker 01:
So, Your Honors, if there are no further questions, I will get back the rest of my time. Thank you for your consideration.

[00:24:00] Speaker 02:
Thank you, Your Honor. Under the standard of review for whether this argument was preserved or not under core photonics, the question is, was this a new argument? That's a de novo question. But in any event, we did say, regardless of the standard of review, this was wrong because, again, the portion of the reply was all quotes from the petition.

[00:24:21] Speaker 02:
Tree Search Engine 74, which is what was used in the 1A part of the petition, was also used in the 1D1E part of the petition to describe the Jang algorithm in its entirety. Those citations are at 106 and 108 of the petition. And Dr. Anderson, who was the expert for Zentian, acknowledged, and this is at 2690, paragraph 47 of that Apple's argument was that TreeSearch 74 was what mapped onto 1D and 1E.

[00:24:58] Speaker 02:
This was not a surprise, and he cited to our experts' deposition for that proposition. And again...

[00:25:07] Speaker 02:
The issue of intercluster memory access, that's misdirection here because that is not claimed.

[00:25:17] Speaker 02:
The circuit does the algorithm, but it doesn't say what components do the algorithm. And it doesn't say that the processors have to do the algorithm. That's unclaimed, right? And there are various ways that this could be done. There could be some other circuit. That was one of the embodiments that's in Figure 2 of the 140 patent that's reproduced.

[00:25:38] Speaker 03:
Some other circuit?

[00:25:40] Speaker 02:
No, some other processor within the circuit. If you look at Figure 2, which is reproduced in the red brief. It's just one of the embodiments of the 140 processor, 140, I'm sorry, patent, claim one. There's a separate processor that kind of combines all of the outputs from the different clustered memories and kind of calculates the pruning threshold. And that's actually dependent claim six in claim one. I'm sorry, dependent claim six in the 140 patent claims that search controller processor.

[00:26:12] Speaker 02:
It's not claimed here. This is a comprising claim, and it leaves open what actually does the work of the comparisons. And so all the boards focus on memory latency and all these problems that it thought would occur because of the need for communication among the clusters. That all goes to an unclaimed element.

[00:26:33] Speaker 00:
Thank you. We thank both sides.