This week in AI: Addressing racism in AI image generators

Keeping up with a rapidly evolving industry like artificial intelligence is a daunting task. So until artificial intelligence can do that for you, here’s a handy roundup of the latest stories in machine learning, as well as notable studies and experiments we didn’t cover ourselves.

On the artificial intelligence front this week, Google suspended the ability of its artificial intelligence chatbot Gemini to generate images of people after some users complained about historical inaccuracies. For example, when asked to depict a “Roman legion”, Gemini showed an anachronistic, cartoonish group of infantrymen of different races, while depicting the “Zulu warriors” as black.

It appears that Google, like some other AI vendors, including OpenAI, has implemented clumsy hardcoding behind the scenes to try to “correct” biases in its models. In response to prompts like “show me only images of women” or “show me only images of men,” Gemini will say no, claiming that such images may “lead to the exclusion and marginalization of other genders.” Gemini is also reluctant to produce images of people identified solely by race – such as “white” or “black” – out of ostensibly concerns about “reducing individuals to their physical characteristics.”

Right-wingers see these vulnerabilities as evidence that tech elites continue a “woke” agenda. But it doesn’t take Occam’s razor to see the less sinister truth: Google has been burned by the bias of its tools before (see: classifying black people as gorillas, mistaking heat guns in black people’s hands for weapons, etc.), Now so desperate to avoid history repeating itself, it projects a less biased world — no matter how wrong — in image-generating models.

In her best-selling book White Fragility, anti-racist educator Robin DiAngelo writes about how the erasure of race (in other words, “color blindness”) leads to systemic Racial power imbalances rather than mitigating or mitigating such imbalances.By claiming to “see no color” or reinforcing the idea that simply acknowledging the struggles of people of other races is enough to label oneself “woke,” people continue DeAngelo said avoiding any substantial protections for the subject would do a disservice.

Google’s heavy-handed handling of Gemini’s race-based prompts does not itself avoid the problem, but is a disingenuous attempt to cover up the model’s worst biases. One might argue (and many do) that these biases should not be ignored or glossed over, but should be addressed within the broader context of the training data they generate—namely society on the World Wide Web.

Yes, the datasets used to train image generators often contain more white people than black people, and yes, images of black people in these datasets reinforce negative stereotypes. This is why image generators sexualize certain women of color, depict white men in positions of authority, and generally favor a wealthy, Western perspective.

Some may argue that there are no winners among AI vendors. Whether they address—or choose not to address—the biases of their models, they will be criticized. Indeed. But I think that either way, these models lack explanation—packaged in a way that minimizes their manifestations of bias.

If AI vendors can address the flaws of their models head-on, with humility and transparency, it will go further than haphazard attempts to “fix” biases that are inherently irreparable. The truth is, we all have biases, so we don’t all treat others the same way. Neither does the model we are building. We’d better admit it.

Here are some other noteworthy AI stories from the past few days:

• Women in Artificial Intelligence: TechCrunch has launched a series highlighting prominent women in artificial intelligence. Read the list here.
• Stable Diffusion v3: Stability AI has announced the launch of Stable Diffusion 3, the latest and most powerful version of the company’s image generation AI model based on a new architecture.
• Chrome gets GenAI: A new Gemini-powered tool in Google Chrome lets users rewrite existing text on the web, or generate entirely new text.
• Darker than ChatGPT: Creative advertising agency McKinney developed a quiz game “Are you darker than ChatGPT?” to reveal artificial intelligence bias.
• Call for legislation: Earlier this week, hundreds of AI celebrities signed an open letter calling for anti-deepfakes legislation in the U.S.
• AI matching: OpenAI has a new customer in Match Group, the owner of apps like Hinge, Tinder and Match, whose employees will use OpenAI’s artificial intelligence technology to complete work-related tasks.
• DeepMind Security: Google’s artificial intelligence research arm, DeepMind, has launched a new organization, Artificial Intelligence Safety and Orchestration, which is made up of existing teams dedicated to artificial intelligence safety but also expanded to include a new group of professional GenAI researchers and engineers.
• Open models: Less than a week after the latest Gemini models, Google has released Gemma, a new family of lightweight open models.
• House Working Group: The U.S. House of Representatives has established an artificial intelligence task force, which, as Devin writes, feels like a gamble after years of indecision with no end in sight.

More machine learning

Artificial intelligence models seem to know a lot, but what do they actually know? Well, what is the answer. But if you phrase the question slightly differently… they do seem to have internalized some “meaning” similar to what humans know. Although no artificial intelligence truly understands what a cat or a dog is, could it encode some sense of similarity in the embeddings of these two words that is different from cat and bottle? Amazon researchers believe so.

Their study compared the “trajectories” of similar but different sentences (e.g. “The dog barked at the thief” and “The thief caused the dog to bark”) with grammatically similar but different sentences (e.g. “The cat slept all day”). and “A girl has been jogging all afternoon.” They found that, despite differences in syntax, what humans perceive as similar is indeed internally perceived as more similar, and vice versa. Well, I find this paragraph a bit confusing, but suffice it to say that the meaning of coding in the LLM seems to be more powerful and complex than expected, without being completely naive.

Researchers at the Ecole Polytechnique Fédérale de Lausanne (EPFL) in Switzerland have found that neural coding could prove useful in prosthetic vision. Artificial retinas and other methods of replacing parts of the human visual system often have very limited resolution due to the limitations of microelectrode arrays. Therefore, no matter how detailed the input image is, it must be transmitted at very low fidelity. But there are many ways to downsample, and the team found that machine learning does a good job at this.

Image Source: EPFL

“We found that if we applied a learning-based approach, we got better results at optimizing sensory encoding. But even more surprising was that when we used an unconstrained neural network, it learned to imitate itself all aspects of retinal processing,” Diego Ghezzi said in a press release. Basically, it does perceptual compression. They tested it on mouse retinas, so it’s not just theoretical.

An intriguing application of computer vision by Stanford University researchers sheds light on a mystery about how children develop drawing skills. The team solicited and analyzed 37,000 drawings drawn by children of various objects and animals, and (based on the children’s responses) analyzed the recognizability of each drawing. Interestingly, it’s not just iconic features like the bunny ears that make the drawing more recognizable to other children.

“The features that caused older children’s drawings to be recognized appear to be driven by more than just a single feature that all older children learn to include in their drawings. What these machine learning systems are dealing with is much more complex. ” said lead researcher Judith Fan.

The chemists (also from Ecole Polytechnique Fédérale de Lausanne) found that LL.M.s were surprisingly good at helping them do their jobs even with minimal training. It’s not just doing a chemical reaction directly, but fine-tuning the work that an individual chemist can’t possibly know all about. For example, there may be hundreds of statements out of thousands of papers about whether high-entropy alloys are single-phase or multi-phase (you don’t have to know what that means – they do). The system (based on GPT-3) can be trained on such yes/no questions and answers and quickly be able to draw inferences from them.

This is not a huge improvement, just more evidence that the LL.M. is a useful tool in this sense. “The point is, it’s as simple as a literature search and applies to many chemistry problems,” said researcher Berend Smit. “Querying the underlying model may become a regular way to start a project.”

Finally, a word of warning from the Berkeley researchers, although now that I’m reading the article again I see that EPFL is also involved in this. Go to Lausanne! The team found that images found through Google were more likely to reinforce gender stereotypes about certain jobs and words than text that mentioned the same thing. And there were more men present in both cases.

Not only that, in one experiment they found that when studying a character, people who viewed images rather than read text more reliably associated those characters with one gender, even days later. “It’s not just about the frequency of gender bias online,” said researcher Douglas Guilbeault. “Part of the story here is that there’s something very sticky, very powerful about the representation of characters in images that text doesn’t have.”

With things like the Google Image Generator Diversity Controversy going on, it’s easy to overlook the established and often proven fact that many AI models’ sources display severe bias, and that bias has a negative impact on people. had a real impact.

Source link

from Tech Empire Solutions https://techempiresolutions.com/this-week-in-ai-addressing-racism-in-ai-image-generators/
via https://techempiresolutions.com/

from Tech Empire Solutions https://techempiresolutions.wordpress.com/2024/02/25/this-week-in-ai-addressing-racism-in-ai-image-generators/
via https://techempiresolutions.com/

from Paxton Willson https://paxtowillson.wordpress.com/2024/02/25/this-week-in-ai-addressing-racism-in-ai-image-generators/
via https://techempiresolutions.com/