When It Comes to Gorillas, Google Photos Remains Blind (wired.com)

Tom Simonite, writing for Wired: In 2015, a black software developer embarrassed Google by tweeting that the company's Photos service had labeled photos of him with a black friend as "gorillas." Google declared itself "appalled and genuinely sorry." An engineer who became the public face of the clean-up operation said the label gorilla would no longer be applied to groups of images, and that Google was "working on longer-term fixes." More than two years later, one of those fixes is erasing gorillas, and some other primates, from the service's lexicon. The awkward workaround illustrates the difficulties Google and other tech companies face in advancing image-recognition technology, which the companies hope to use in self-driving cars, personal assistants, and other products. WIRED tested Google Photos using a collection of 40,000 images well-stocked with animals. It performed impressively at finding many creatures, including pandas and poodles. But the service reported "no results" for the search terms "gorilla," "chimp," "chimpanzee," and "monkey."

Comments

Re:How does that work in practice?

[Solandri]

You're trying to be funny, but I think this highlights part of the problem. It's not necessarily Google's image recognition software, it's also poor photographs. A self-driving car (if it needed to distinguish between gorillas and black people) wouldn't have as much problem because it can adjust the camera's exposure to where it can see enough detail to distinguish the two. But a lot of photos of black people are taken with the wrong exposure, and the black skin tones end up crushed down into just a few discrete color values near black, or clipped to 0. The AI then ends up trying to distinguish one black humanoid-shaped blob from another.

When I shot weddings on film, I had to use special low contrast film (had a larger dynamic range). That was the only way to retain detail in both the bride's white dress and the groom's black tuxedo. And even then, if the wedding was held in sunlight a lot of the detail might still be unrecoverable. Modern cameras are getting around the problem with automatic HDR photo mode (takes two photos at different exposures to preserve detail in both the highlights and shadows, then combines them nonlinearly). But there's still a huge library of badly-exposed photos out there (from the pre-HDR days and being added to by current photographers taking simple snapshots without really caring about exposure), just waiting to trip up any image recognition AI.

White skin tends to be slightly brighter than the average background, while black skin tends to be much darker. So to properly expose a portrait of a black person, you have to either make sure their skin dominates the camera's auto-exposure algorithm, and not the background or their clothing. Or put additional light specifically on their face (e.g. fill flash) so it's not so dark relative to the background. A professional photographer knows this. The average person taking a snapshot, and the auto-exposure algorithm in their camera, does not.

Re:Black Panthers was perfectly acceptable

[rahvin112]

Either you are being deliberately obtuse or you're an idiot.

Gorilla is offensive for the same reason other terms used derogatorily are. It was frequently used as a term of offense during the slave trade and jim crow. There are references going back to the 1600's when the slave trade started referring to humans with dark skin as gorilla's or apes.

But go ahead and think it's not a big deal because you're an idiot, you'd think differently if someone had used the term to refer to you as sub-human.

Re:Black Panthers was perfectly acceptable

[rahvin112]

That don't apply anymore? Don't be a fucking idiot. Gorilla is a term as frequently used as any other racial slur. It's used just as often today as all the other terms as it was when it first started to be used.

Re:People look like apes, black people more so

[AmiMoJo]

Bollocks. You have zero evidence that is true, and in fact no one seems to have been fired and Google assigned engineering resources to finding a solution.

It's hard because the AI is simplistic. It's actually a good test case.