The Worlds I See by Fei-Fei Li

Author:Fei-Fei Li
Language: eng
Format: epub
Publisher: Flatiron Books

The following months fell into a rhythm, though not a particularly graceful one. ImageNet was a wild animal that refused to be tamed, lashing out every time we got too close. We kept at it, scoring our share of victories—little ones, at least—alongside accumulating scrapes and bruises. But every time we thought we’d finally backed it into a corner, it summoned a deeper, more guttural roar, and sent us scurrying.

Luckily for me, Jia was the kind of partner who reacted to a frustrating problem by thinking harder. Human participation was the costliest part of our process, both in terms of time and money, and that’s where he began his counterattack: making it his personal mission to reduce that cost to the absolute minimum. For example, when one of our labelers set about curating a collection of photos for a particular category, say, “Pembroke Welsh corgi,” we initially expected that each step would be carried out manually: typing the query into a search engine like Google Images, combing through the results to find clear examples, applying the label to each, then placing the final selections in the proper directory. But most of those steps didn’t require human intelligence.

The first thing Jia automated was the downloading phase, writing a program that submitted each WordNet category to an image search engine as our labelers had been doing. But since search engines are meant for human users, not machines, they don’t return a set of images directly; instead, they present a web page that organizes the resulting hits as a scrolling grid of thumbnails, the source code of which Jia’s program would then parse to extract links to the full-sized images themselves. It was a messy solution, but it gave us the ability to download candidate images at maximum speed, day and night, for as long as we wanted—months, if necessary. And the resulting images were automatically organized on our own machines.

Our repository began filling up like magic. Granted, the wide net we’d cast was catching a fair amount of junk—low-quality photos, clip art, and the like—but we were accumulating plenty of good stuff, too. Somewhere in our network of rapidly filling hard drives, the first glimpses of that mosaic were coming together—a crude but authentic depiction of the entire visual world. At least, it was for a while.

“Uh-oh,” I heard Jia say from across the lab.

“What’s the matter?”

“Looks like we’ve hit a bit of a speed bump. Uh … yep. Google’s banned us.”

“What? Banned? Why?”

“Evidently they cap the number of requests a single user can submit in a specific period. About a thousand or so, from what I can tell.”

“How long is this period?”

“Twenty-four hours. It resets at midnight. That’s the good news.”

“Okay, how fast are we burning through the daily allotment?”

“Well, that’s the bad news.” Jia pulled up the log file and did some mental arithmetic. “About nine minutes.”

Oof.

The growth of the repository stalled. And that wasn’t our only problem. The pipeline was cartoonishly lopsided; our collection of raw images was exploding,

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.