They both knew enough of what Corrigan was talking about to make questions unnecessary. All he needed was a cue. Pinder nodded. “Go on, Joe. How?”
Corrigan moistened his lips. “The top-down, analytical approach doesn’t work. Everyone in the field agrees. The only way it’s going to happen is by getting some kind of initially simple system to evolve.”
“Which has been tried in enough places too,” Pinder observed. “And the results have all been equally modest, to say the least.”
“Agreed. But they’ve all been tries at equipping computers with sensory apparatuses like TV cameras, arms, legs, and wheels, and letting them loose to explore some kind of environment. But you don’t realize how good biological nervous systems are until you try copying them. They were shaped by a billion years of evolution to interact with the real world. Computers weren’t.”
Which exhausted what everyone in the trade knew were the two acknowledged theoretical approaches. “So are you saying you know another way?” Pinder asked.
“Yes, I think so.”
“What?”
“Computers do interact extremely well with their own, internal worlds. . . . So what you do is, invert the conventional approach.” Corrigan spread his hands. “If training a machine intelligence in our world isn’t effective, let’s try doing it the other way around: by going into its world and doing it there.”
Pinder frowned. “Sorry, I’m not quite with you, Joe. Doing what, exactly? Where?”
“People interfacing via EVIE interact with a machine-created version of the real world through the surrogates that they control. But the machine could also put pseudopeople of its own in there too—`animations.’ You design the system to be goal-directed to make the behavior of its animations converge to that of the real-people surrogates.”
Pinder sat back, seeing the implication at once and staring at Corrigan thoughtfully. “So its success would be measured through a kind of Turing test,” he said.
“Yes, exactly.”
“This is certainly a new one on me, Joe. I’ve never heard the like of it.”
“What do you think?”
“It’s intriguing.”
Corrigan could see that he was making an impression and pursued his point further. “The system wouldn’t need to know why the individuals that it was trying to imitate were doing whatever they did. Its brief would be simply to make its animations behave similarly, which it could accomplish from external observables. And that’s what’s different about this approach. In the past, we’ve always tried to press into service existing processing methods and associative structures—tools that were developed for other purposes. Well, very possibly they’re inherently unsuitable for this kind of job and can never work. But the way I’m talking about, the system will be free to create its own organization of associations and linkages in a way that’s appropriate to its goals.”
“Information-processing architecture is appropriate to what information-processing systems do. Whatever it is that has evolved inside cerebral cortexes is appropriate to what cerebral cortexes do,” Pinder summarized.
“That’s it. And we don’t need to know in advance what the final organization will be, any more than the first protoplasm needed to know the wiring for a mammalian brain. The system would learn the way children do: by trying to imitate `adults’ who already understand the way the world works, and making its own connections and associations accordingly.
“And we’ve got all the pieces needed to do it. Pinocchio provides the basics of a suitable vehicle for driving both the surrogates and the animations. EVIE, with the all-neural package that we’re talking about for COSMOS, gives us a mechanism for coupling in the surrogates. A multitasking expansion of Jenny Leddell’s Perseus system from MIT could drive the animations.”
Corrigan judged this a good place to stop at for a response, and waited. Pinder stroked his chin and stared down at the desk. What Corrigan was proposing was clear enough. He was searching for the flaws. Finally he looked up.
“A world to support that kind of evolution needs to be context-rich,” he said, meaning the degree of detail and its variability that the system would have to support. “The look-ahead for sudden context changes and recomputing SDVs still hasn’t been solved satisfactorily. And it would get a hell of a lot worse with this.”