"Marcel? How would you like a chance to redeem yourself?"
"Well, as I understand it—"
"Oh, don't give us that, Marcel. Give us the facts."
"As I understand it, you present the net with a pattern of input. This signal pattern flows from neurode to neurode along branching, variously weighted connection paths. If the sum of inputs on the receiving neurode exceeds its signal threshold, it, too, fires and passes along more signal. Spreading activation, it's called."
I looked at Lentz to see if I had it right so far. He had his hands together, fingers to lips. And he was smirking.
"The signal pattern spreads through the net from layer to layer. A final response collects at the output layer. The net then compares this output to the desired output presented by the trainer. If the two differ, the net propagates the error backward through the net to the input layer, adjusting the weights of each connection that contributed to the error."
"Bra-vo, Marcel. Who's your teacher? Now, the question is: Which would you rather be able to do? Explicate that process or interpret Donne's Holy Sonnet?"
"I don't think that's the question," Diana sighed. "The question is whether back-propagation violates the directionality of the axon-to-dendrite signal."
Lentz pulled his head back on its stalk. He arched his eyebrows. "What have we here? The dabbler's expertise!"
Diana looked as if he'd just slapped her. Her lip shook. She would have stormed off if Harold hadn't restrained her.
"Philip!" Plover frothed. "Just shut up for a second, will you? You fault us for taking issue from a position of ignorance. But we're challenging you to provide the refutation. Isn't that better than reflex rejection? Torches and scythes?"
"Torches and scythes are ever so much more picturesque."
"Oh, for the love of… I give up." Plover traced a shorthand of his bear-paw exasperation.
"Uh-oh. About to invoke the deity. We're in trouble now. Whatever happened to radical skepticism in religious matters?"
"Philip, you. ." Plover tripped on his tongue in his disgust. "Why bother publishing your results, if you don't want people to follow you?"
"I get nervous when I'm followed too closely." Lentz inserted a fry into his mouth like a young child hammering blocks into a toy workbench.
"So you can't be bothered with anyone who isn't already in the inner sanctum."
"That's not true. Why, I've taken Marcel here and have already turned him from near-ignoramus into marginally literate page boy."
" 'Sweets grown common lose their dear delight.' Is that it? Lentz, that's the most elitist thing I've ever heard anyone on public funding say."
"Marcel, tag that line for me, will you?"
Citing seemed the quickest way out of trouble. "I think it's Shakespeare."
"Shakespeare's not elitist?" Lentz countered. "In any case, most of my grants are corporate, these days. By the way, Harold. 'Most elitist' is not up to your usual elegant prose."
"When you boys are finished," Diana murmured, "I wouldn't mind an answer."
Lentz cackled. "All right, Marcel. Time to straighten out the masses again."
"I guess the idea is that backward error propagation may resemble higher-order brain processes. True, individual nerve pathways are one-way. But neuronal paths as a whole do connect portions of the brain in two directions."
"I get it," Diana said. "Feedback signals from muscle tissue to the primary motor cortex, and the like?"
Lentz snorted. "This is just Marcel's guess, you understand."
I wondered if I would be able to work with the man a whole ten months.
Diana's face wrinkled. "But it's not quite the same thing, is it? Is presynaptic Hebbian change the same thing as…?"
But Lentz had stopped listening. He was peering into his juice bottle, preoccupied. Almost consumed by distraction. "Nothing," he declared. "Nothing is the same as anything else."
By whatever mechanism, that lunch is set into my cortex. I can take out the tape and view it at will. From this distance, its edges are ridiculously sharp, the definition absurd. These frames extend a longer clip, animated by the persistence of vision.
Diana later told me how memories are laid down when the thumb-sized basal forebrain bathes the hippocampus in acetylcholine. The chemical somehow changes a synapse's shape, she claimed, altering the connections between cells. And a key flooder of the chemical pathways is fear.
That lunch is burned into my memory because I sat through it terrified. Lentz scared me. Harold's hurt, Diana's dismay triggered my impulse to fly. I was afraid of everything these three might say to each other. Spooked by the open rupture begging to take place. Every word of anger was my fault.
For a long time, I thought that what frightened me was the prospect of failure. My biggest anxiety seemed to be that we'd pour ourselves into this folly and never get it to do more than gibber. That a previously productive researcher would take me on a ten-month wild-goose chase.
It's taken me this long—this long — to see. The fear that laid down that indelible trace was the same one I've nursed since boyhood. The fear that we might realize our dreams.
Lentz built Implementation A more for my education than as a prototype with any real pretensions. The beast wasn't winning any beauty prizes, either in looks or in conception. Lentz cobbled up a card cage to the back plane of your basic vanilla workstation. A few I/O devices, registered antiques all, hung off the ports, windows to this poor contraption's soul.
"The super-rich always drive beaten-up old Chevies," Lentz assured me. "You have to grasp the inverse snob value. Always dazzle your audience with misdirection. If we get Three Blind Mice' to come out of this, it'll knock them dead."
However humble, the rig gave us an entree. A workable learning algorithm can run on any platform. The brain, Lentz had it, was itself just a glorified, fudged-up Turing machine. Our cerebrum-to-be had no neurons, per se. No axons or dendrites. No synaptic connections. All these structures hid in simulation, dummied up in the standard linear memory array. The bare troika of Boolean operators brought them into metaphorical being. We used algorithms to imitate a non-algorithmic system. Implementation A was a ghostly hologram. It froze our words the way a scribbled shopping list, falling from a book where it has spent a misplaced lifetime, revives that longhand you thought you'd live alongside forever.
Lentz did a good job of making the hardware transparent to me. He hooked up topologies the culmination of a decade or more of tinkering. He explained every link in the process. But I was exhausted, wiped out from my own recent work. I tried for the gist and took the rest on faith.
The gist consisted of vectors. A stimulus vector, converted by the net's self-reorganization into a response vector. We started with a three-deep array of neurodes, enough for a test start. Each field was the size of the net that had learned to pronounce English. Implementation A would be spared this task. Lentz wired it to a canned speech synthesis routine. We worked at the level not of phonemes but of whole words.
At first I typed input into the system. My text ran through a huge lookup table — Lentz's laborious list of the 50,000 most common English words, arranged in order of rarity. Each word got a number, like a runner in a triathlon. These numbers in turn plunged headfirst into our randomly weighted terrain.
The ersatz brain cells juggled the pattern, until those in the output layer replied in noise. The box answered back each time I typed to it. But it answered gibberish. Like any nursling's, its voice was awash in gabble.
If I kept at the same stimulus, however, the output organized. Activated paths strengthened; inert routes atrophied away. The blathering gave way to consistent conversational response. What those responses were, however, neither Lentz nor I could say.