“In fact, we will provide each of you with an Imperial Finding those charters remain in effect and binding on the Throne.”
“That is most welcome news, Your Majesties,” Khan said. “We have worked together, the five of us and the Empire, for a third of a millennium, and we hope our relationship lasts centuries more.”
“OK, that’s everybody?” Burke asked.
“That’s everybody,” Ardmore said. “And now we’ve normalized relations between the Throne and almost all the anti-Throne forces. Only one enemy left.”
“Yes. The real enemy. I wonder how the Zoo is doing?”
“I understand they’re close.”
Back to the Zoo
While Ardmore and Burke worked on the sector governors and the royal heirs, the Zoo was working on the alias problem. They had had some success with tracking real names to alias names, but they wanted more.
“What if we map out all the conjugate sounds? You know, like d and t, and g and k, and all the diphthongs. Map it all out and program the equivalencies,” Rick Pender said.
“That might work,” Lucia Martelli said.
“It sounds like a lot of work,” Matthew Houseman said.
“I can’t help thinking we’re looking at this the wrong way,” Lois Costas said.
“What do you mean?” Denise Coutard said.
“Well, we have all this computer power, right?” Costas said. “And we’re thinking of ways to do the computation differently. Do a lot of computer programming to try and get better solutions. I’m thinking that’s not the way to go.”
“Where’re you going with this, Lois?” Houseman asked.
“I think we ought to spin up a neural net, give the computer all the right and wrong answers so far, and let it figure out how to do the compares to get our answers,” Costas said. “Then turn it loose on the whole data set.”
“That’s days of computer time, Lois. Maybe weeks. I’m not sure we can do that.”
Wang Minwei disappeared, then popped back into existence a minute later.
“Olivia says this is a high-priority item,” Wang said. “The highest. If we have a way to do it better, we can have all the computer time we need.”
“Well, I think the neural net is the way to go,” Costas said. “It’s a resource hog, but it’s going to get the best final answer.”
“Let’s start parameterizing the net and getting the initial feed set up then,” Houseman said.
“I’ll warn the computation center,” Wang said and disappeared.
Wang’s warning caused more than a little excitement in the Imperial Navy Computation Center.
“Hey, Manny, we gotta clear the decks. We got incoming,” Delores Clark Foley said.
“Why, Dee? What’s going on?” Manuel Delgado asked.
“I just got a warning. The Zoo wants to spin up a neural net. A big one.”
“What?”
“Yeah,” Foley said. “They’re gonna teach the computer English or something.”
“No shit. Do they have the priority for a job like that?”
“It’s an Imperial priority, Manny. And I would just as soon not have the Empress call me up to explain it to me. So fire up all the off-line capacity, kick off everybody who isn’t at priority one or two, and shunt those who are off onto one of the standbys. Clear all the main grid. Of everything.”
“All right, Dee. You got it. Gonna take me fifteen, twenty minutes.”
“That’s fine, Manny. It’s a thirty-minute warning.”
Twenty minutes later, the gigantic computation engine of the Imperial Navy’s battlespace integration and visualization simulator was standing idle, waiting for the Zoo to spin up the neural net.
“We’re ready to go?” Houseman asked.
“Yeah, we’re set with the neural net parameters,” Costas said.
“How bad is it?”
“You don’t want to know.”
Houseman gave her a sharp glance, and she shrugged.
“It’s big, OK? We’ll see if it’s too big soon enough.”
“How about the feed? We got all the right-answer/wrong-answer data ready to go?”
“Yeah we’re set,” Pender said.
“OK, Lois. Spin up the net.”
“Here it comes,” Foley said, watching the performance monitors.
“Oh my God. Look at it,” Delgado said.
“We have everything on line?”
“Yes, that’s all we got.”
“It’s still spinning,” Foley said. “I wonder how big this is going to be.”
“It’s starting to top out now.”
“Where are we?”
“Seventy-five percent,” Delgado said.
“Is it going to fit?”
“We should see soon.”
“There,” Foley said. “It’s rolling out the edges.”
“Yeah. She’s up. The whole thing.”
“What are we at now?”
“Ninety-three percent,” Delgado said.
“Holy shit.”
“The net has spun up,” Costas said.
“Did it fit?” Houseman asked.
“Yeah. We’re at ninety-three percent.”
“How much of the battlespace simulator is on-line?”
“All of it,” Costas said.
“We’ve got the whole thing? Everything?”
“Yeah. Ninety-three percent of everything they got over there. They dusted off all the electrons they have.”
“Wow,” Houseman said. “All right, Rick, feed the answer data.”
“In progress,” Pender said.
“Are you feeding the phonetic versions?”
“No. The plain-text versions. The phonetic data is part of the resource field.”
“Are you sure about doing it that way?” Houseman asked.
“Might as well. Who says everybody pronounces everything the same? Ever talk to someone from Terre Autre?”
Houseman nodded. It was more computer power, but what the hell? If it worked, it would be glorious.
“We have all the data in,” Pender said.
“OK, Lois. Kick it off,” Houseman said.
“Loading initial seed....
“Seed initiated....
“First pass initiating....
“First pass under way....
“OK, Matt. It’s running,” Costas said.
“Well, we might as well all go get some food and some sleep,” Houseman said. “This is gonna take a while. Whaddya think, Lois? Twelve hours?”
“Might as well make it sixteen, Matt. This is gonna be a while.”
What the computer did running the neural net is run the plain-text names against a pattern-matching algorithm. The initial algorithm was a randomly-seeded decision matrix. The data with known match/no-match answers was run through the algorithm, and the results recorded. All the hundred thousand pairs of names they had looked at were run through, and the computer scored its results against the correct answers.
The system then made a random set of adjustments to the decision matrix, and ran all the data through the algorithm again, and scored its results. It compared the scores. Did the random changes help or hurt? If they helped, it made a random set of adjustments to the second decision matrix. If they hurt, it backed up and made a different set of random adjustments to the first decision matrix.
The process was not unlike watching an infant trying to fit differently shaped blocks through variously shaped holes. The computer kept trying things, looking for the decision matrix that fit the problem, giving the same results as the human decision makers.
Like the infant, as it manipulated the data, it was learning.
And, like the infant, it was a slow, iterative process.
The team came back into the Zoo after dinner, a normal evening at home, a full night’s sleep, and even breakfast.
“How’s it doing, Lois? Are we getting anywhere?” Houseman asked.
“Yes. It looks like it’s right about eighty p
ercent of the time now, with the current decision matrix.”
“OK, everybody,” Houseman said. “Let’s take a look at the answers where the current decision matrix and our answers disagree, and look for any mistakes we might have made.”
This was always the problem with these sorts of problems. With computers, it was always garbage-in gives you garbage-out. The version of it that applied to this sort of problem is that wrong answers in the answer key the computer was trying to emulate led the decision matrix down the wrong path, and could keep it from closing in on an acceptable error rate.
The team spent the next eight hours poring over the name pairs where the current computer assessment and the original human answer key varied. Looking for their own errors. During the course of the day, they found seven hundred errors in the original hundred thousand answers, an error rate under one percent. They hadn’t done badly, but those errors could throw off the decision matrix badly.
“All right, I guess it’s another quiet evening at home everybody,” Houseman said. “See you all tomorrow.”
“Are they running data now?” Foley asked.
“No, they’re still in the learning process,” Delgado answered.
“How long has it been running?”
“Coming up on forty hours now.”
“CPU utilization?” Foley asked.
“It’s been hard on ninety-nine percent the whole time.”
“How’s our cooling doing?”
“We were running temperatures higher than normal for a while, but we’re good at the moment,” Delgado said. “I brought more refrigeration on line about eight hours back.”
“How’d you manage that?”
“I shut off the air conditioning in the rest of the building.”
“I wondered what was going on,” Foley said. “It’s getting pretty hot in my office.”
“Can’t be helped. I can’t spare the capacity right now.”
“How are we doing today, Lois?” Houseman asked.
“It looks like it’s getting close. Ninety-eight percent or so.”
“Let’s look at the differences again, everybody. See if we missed any errors last time we looked.”
The team found another fifty errors. Part of that was because every time they looked at the data, a different person was looking at each specific name pair, so they would be unlikely to make the same mistake. Part of it was they were concentrating more as the number of differences diminished.
Costa watched the algorithm throughout the day. As they entered the corrections in the human-generated scorecard, the neural net was making smaller and smaller adjustments to the decision matrix.
“It’s starting to close, Matt,” Costas said.
“Are we ready to go with the problem data, Rick?”
“Yeah. You sure you want to run it all, Matt?”
“Yep. All the descendant names against all the alias account names, beginning with the current generation and working back.”
“That’s an m times n problem. It’s gonna take a while.”
“I don’t care, Rick. We’ll take it one generation at a time.”
“Right you are. We’re ready.”
It was late in the day when Costas announced success.
“It’s closed on ninety-nine-point-nine-plus. It’s tried to modify the matrix for better results unsuccessfully over a hundred times.”
“All right,” Houseman said. “Let’s stop the learning process and start it working on the problem data.”
“Right,” Costas said.
“Whoa! Look at it go,” Pender said.
“Now that it has the decision matrix, it’s gonna scream,” Costas said. “It’s a massively parallel problem, but it’s a massively parallel machine.”
“Well, with six hundred million people in the current generation and millions of alias accounts, it’s got trillions of compares to do,” Pender said. “It may be fast, but it’s still going to take a while.”
“Yeah, but we have answers coming out already,” Costas said.
“Let’s spot-check those,” Houseman said. “Do they look like the machine is finding good matches?”
Costas, Pender, and the rest of the team pored over the results. The vast majority of the compares were no-match, as you would expect, but the matches were what they were looking for.
“Yeah, these look good, Matt.”
“Excellent,” Houseman said.
He looked at Costas.
“That was a really good idea, Lois.”
“They’re running data now,” Delgado said.
“Good,” Foley said. “They’ll be done soon and we can have our air conditioning back in the rest of the building.”
“I wouldn’t hold my breath.”
“Why not?”
“It’s two really big data sets, and it’s an m times n problem,” Delgado said.
“Oh, great. Is it going to be done before I retire?”
Delgado laughed.
“That’s sort of up in the air at the moment.”
Every day, the team worked on the output from the learned algorithm, reviewing the machine’s decisions and marking the positive matches either likely or less likely, but keeping both types in the dataset. They were quickly overwhelmed, and the rest of the Zoo jumped on the problem. They gradually caught up, and, as results were checked, they were posted to the Zoo’s bulletin board.
As results were posted, Olivia Darden transmitted them to Lina Schneider in the Imperial Investigations Office.
“What do we do with this new data?” asked Stanley Nowak, Schneider’s assistant department head and chief investigator.
“I think we want to do a heat map sort of thing. So display all the ‘likely’ ones in red and the ‘less likely’ ones in orange. Then the parent and grandparent accounts inherit the hotter of the two.”
“OK, so if a parent account has one likely and three less likely, we’re still going to mark it as red, and the same with any grandparents?”
“Yeah, I think that’s the way to approach it,” Schneider said.
“OK. There’s quite a bit of data there.”
“Oh, this is just the first tranche of output data. It’s going to continue to come in for a while.”
“Really?” Nowak asked.
“Yeah, they threw the kitchen sink in there. They’re processing all the descendants against all the alias accounts, in the whole Empire.”
“Wow. Well, that explains why it’s taking a while. No matter what computer they’re using.”
Schneider sat in the viewing room and watched the data grow in the display of all the alias accounts in the Imperial Bank. There were so many alias accounts, the entire map of them looked like a huge cloud of gnats. Initially, the new data coming in were isolated red and orange dots sprinkled across the map.
As the incoming data grew, and the parent and grandparent accounts inherited the child account status, blots here and there in the map turned red and orange. Those blots grew as the data kept coming in, and some of them linked to each other. The map reorganized itself to keep connected blots in proximity to each other.
Schneider watched the red stain growing in the display over the course of the next week.
It took twelve days to run the entire data set through the big computation engine. People in the computation center took to working from home as a bit of a heat wave hit Imperial City and their offices became unbearable without any air conditioning in the office spaces of the building.
“Well, that’s it,” Delgado said. “They’re done.”
“Finally,” Foley said. “Does that mean we can have our air conditioning back?”
“Yeah. Nothing left to do now but the clean-up. They want us to hang onto the decision matrix and the output files. That’s no problem. Then I need to open the system back up to the user base. Some people are getting pretty pushy about it.”
“What are you telling them? Anything?”
“Nope. Secret job. Imperial Pri
ority. Take it up with Their Majesties.”
“OK. Well, that should shut them up.”
“Oh, it does.”
Delgado looked at Foley for a long moment.
“You know, Dee, we did that big upgrade to the main grid over the last couple years, right?”
“Yeah. Sure. So what?”
“And it’s the biggest machine anywhere.”
“Yeah? And?” Foley asked.
“Well, it took the machine four days to learn their decision matrix and another twelve days to run the dataset. Call it sixteen DAYS of computer time at next to a hundred percent of utilization.”
“I still don’t get your point, Manny.”
“Dee,” Delgado said, “what you just witnessed was the largest computer problem ever run, by anyone, ever, in history.”
“No shit.”
“Yeah. And we can’t tell anyone about it.”
“Well, we can tell the Zoo,” Foley said.
It took the Zoo another day to catch up to the end of the data output from the decision matrix and post their final results.
“All right, everybody,” Houseman said. “Nice job. Spin-down time. See you all next Monday.”
“It’s only Tuesday now, Matt,” Pender said.
“Have a nice weekend,” Houseman said with a smile.
Schneider was looking at the investigation map once all the data had been entered. She was focused in on one area of the map, tracing some connections. Nowak joined her.
“Well, all the data is in now. What does our map look like?” Nowak asked.
“You sure you’re ready for this?” Schneider asked. “I’ve been watching it develop the last two weeks.”
“Why do you ask? Is there something wrong with it? I haven’t looked at it at all, because I’ve been working with the data people.”
“Brace yourself.”
Schneider dialed the view out so the entire map was displayed. An ugly red stain spread across almost half the map.
EMPIRE: Resistance Page 25
