out of the way, so we can get to work.”
“Okay, for starters, why virtualize?” she asked. “You don’t need virtual machines when you’re the only user.”
“Think of genetic algorithms,” I said. “You ask a question, and, like a GA, Frankenstein starts from lots of different random assumptions simultaneously to find a logical answer. He searches all of them in parallel, assigning a changing level of resources to each solution path based on ongoing evolutionary fitness, until one of those paths reaches an answer he likes. So why should I reinvent the wheel when I can use off-the-shelf virtualization to manage that part?”
“But there’s something weird about it, Trevor. Frankenstein’s different virtual machines can come up with multiple solutions to a problem. Ask it the same question, and it won’t necessarily answer the same way each time.”
I shrugged. “Do you?”
“I’m not a computer,” she said. “This is going to take a little getting used to, that’s all. But I’m wondering, what happens to all the other virtual machines after Frankenstein answers the question?”
“He leaves them running in the background. Most will converge to a suboptimal solution and shut down by themselves. But sometimes they keep going and eventually come up with a different answer that he likes even better.”
Cassie gave an incredulous laugh. “So Frankenstein not only gives unpredictable answers, but it can change its mind after the fact, too?”
“It’s part of how he learns.”
“About that,” she said. “When I first heard Frankenstein speaking to you, I expected you’d built a layered expert-system inference engine like DeepQA. You know, what IBM’s Watson supercomputer used—”
“—to win the Jeopardy world championship back in 2011?” I snorted. “As if beating humans in a game show was supposed to be difficult or something?”
“Winning Jeopardy was only a demo,” she said. “Since then, they’ve retrained Watson on oncology and radiology data, feeding it CAT scans, patient histories, and MRIs. IBM installed it at Sloan-Kettering, where it’s more or less serving as a consulting physician now, recommending diagnosis and treatment plans for cancer patients.”
“Watson’s a toy,” I said. “It maxes out at what? Eighty teraflops? Frankenstein’s got seven hundred times more processing power.”
Her eyebrows went up in surprise. “That’d make Frankenstein the…”
I nodded.
“But we’re talking about software, not hardware,” she said. “Your learning algorithms aren’t expert-system inference engines like Watson’s, and they aren’t neural network-based either.”
“Both approaches are a waste of time,” I said. “An expert system is a closed-loop dead end, because the inference engine is only as effective as the human-curated dataset you feed it. And neural networks—why use modern semiconductor technology to model an inefficient biological kluge that’s five hundred million years old? Imitating organic brain structure is flat-out stupid—like welding fake trees out of steel so you can chop them down and build yourself a log cabin. Much smarter and cleaner to use hidden hierarchical Markov models and layer them deep, like I did.”
She shook her head. “But this way, we—”
“With the kind of processing power Frankenstein has, I can let him do most of the work. I just give him a bunch of different algorithmic building blocks and data taxonomies and let him figure out the optimal way to use them.”
“So you’re cheating.”
I laughed. “No, but I’m smart enough to recognize a job a computer can do better than I can.”
“But don’t you see the problem? This way, we’ll never really understand exactly how the software works—why it answers the way it does.”
“We don’t need to,” I said. “Don’t get confused about the goal of our research here. Frankenstein’s only a tool. Its job is to help us figure out why people answer the
“Okay, for starters, why virtualize?” she asked. “You don’t need virtual machines when you’re the only user.”
“Think of genetic algorithms,” I said. “You ask a question, and, like a GA, Frankenstein starts from lots of different random assumptions simultaneously to find a logical answer. He searches all of them in parallel, assigning a changing level of resources to each solution path based on ongoing evolutionary fitness, until one of those paths reaches an answer he likes. So why should I reinvent the wheel when I can use off-the-shelf virtualization to manage that part?”
“But there’s something weird about it, Trevor. Frankenstein’s different virtual machines can come up with multiple solutions to a problem. Ask it the same question, and it won’t necessarily answer the same way each time.”
I shrugged. “Do you?”
“I’m not a computer,” she said. “This is going to take a little getting used to, that’s all. But I’m wondering, what happens to all the other virtual machines after Frankenstein answers the question?”
“He leaves them running in the background. Most will converge to a suboptimal solution and shut down by themselves. But sometimes they keep going and eventually come up with a different answer that he likes even better.”
Cassie gave an incredulous laugh. “So Frankenstein not only gives unpredictable answers, but it can change its mind after the fact, too?”
“It’s part of how he learns.”
“About that,” she said. “When I first heard Frankenstein speaking to you, I expected you’d built a layered expert-system inference engine like DeepQA. You know, what IBM’s Watson supercomputer used—”
“—to win the Jeopardy world championship back in 2011?” I snorted. “As if beating humans in a game show was supposed to be difficult or something?”
“Winning Jeopardy was only a demo,” she said. “Since then, they’ve retrained Watson on oncology and radiology data, feeding it CAT scans, patient histories, and MRIs. IBM installed it at Sloan-Kettering, where it’s more or less serving as a consulting physician now, recommending diagnosis and treatment plans for cancer patients.”
“Watson’s a toy,” I said. “It maxes out at what? Eighty teraflops? Frankenstein’s got seven hundred times more processing power.”
Her eyebrows went up in surprise. “That’d make Frankenstein the…”
I nodded.
“But we’re talking about software, not hardware,” she said. “Your learning algorithms aren’t expert-system inference engines like Watson’s, and they aren’t neural network-based either.”
“Both approaches are a waste of time,” I said. “An expert system is a closed-loop dead end, because the inference engine is only as effective as the human-curated dataset you feed it. And neural networks—why use modern semiconductor technology to model an inefficient biological kluge that’s five hundred million years old? Imitating organic brain structure is flat-out stupid—like welding fake trees out of steel so you can chop them down and build yourself a log cabin. Much smarter and cleaner to use hidden hierarchical Markov models and layer them deep, like I did.”
She shook her head. “But this way, we—”
“With the kind of processing power Frankenstein has, I can let him do most of the work. I just give him a bunch of different algorithmic building blocks and data taxonomies and let him figure out the optimal way to use them.”
“So you’re cheating.”
I laughed. “No, but I’m smart enough to recognize a job a computer can do better than I can.”
“But don’t you see the problem? This way, we’ll never really understand exactly how the software works—why it answers the way it does.”
“We don’t need to,” I said. “Don’t get confused about the goal of our research here. Frankenstein’s only a tool. Its job is to help us figure out why people answer the
Similar Books
