AI Evolution: The Never-Ending Race to Keep Up
The AI industry moves at an unprecedented pace, making it a challenge for even the most dedicated individuals to stay in the know. Until an AI is able to take over this daunting task, here is a convenient compilation of recent stories in the world of machine learning, including notable research and experiments that may have gone unnoticed.
This week in AI, Google made headlines for pausing its AI chatbot, Gemini, from generating images of people after receiving criticism for historical inaccuracies. For instance, when asked to generate an image of “a Roman legion,” Gemini would produce a cartoonish and anachronistic group of racially diverse foot soldiers, while images of “Zulu warriors” would be predominantly Black.
Told to depict “a Roman legion,” for instance, Gemini would show an anachronistic, cartoonish group of racially diverse foot soldiers while rendering “Zulu warriors” as Black.
It seems that Google, among other AI vendors like OpenAI, had implemented clumsy hardcoding in an attempt to “correct” for biases in their models. However, when prompted to show images of only specific genders or races, Gemini would refuse, citing concerns about “exclusion and marginalization.” Furthermore, the chatbot was hesitant to generate images of people identified solely by their races, claiming it would reduce individuals to their physical characteristics.
Gemini was also loath to generate images of people identified solely by their race – e.g. “white people” or “black people” – out of ostensible concern for “reducing individuals to their physical characteristics.”
Some have accused Google, and the tech elite in general, of perpetuating a “woke” agenda with these biases. However, it is likely that Google is trying to avoid past mistakes, such as categorizing Black men as gorillas or mistaking thermal guns in Black people’s hands for weapons. In an effort to prevent these biases from repeating, Google is manifesting a less biased world in its image-generating models, albeit erroneously.
The Harm of “Color Blindness”
In her best-selling book “White Fragility,” anti-racist educator Robin DiAngelo argues that the concept of “color blindness” only perpetuates systemic power imbalances rather than alleviating them. By avoiding conversations about race or reducing individuals to their physical characteristics, people actually contribute to the harm and continue to uphold these imbalances.
By purporting to “not see color” or reinforcing the notion that simply acknowledging the struggle of people of other races is sufficient to label oneself “woke,” people perpetuate harm by avoiding any substantive conversation on the topic.
The Irreconcilable Issue of Bias in AI Models
Google’s handling of race-based prompts in Gemini did not truly address the issue at hand. Instead, it disingenuously attempted to hide the model’s biases. It’s clear that these biases cannot be ignored or brushed over, but must be examined within the broader context of society and the world wide web from which the training data arises.
Yes, most image-generating data sets contain more white people than Black people, and the images of Black people in these data sets often perpetuate negative stereotypes. This is why certain women of color are sexualized, white men are depicted in positions of authority, and wealthy Western perspectives are favored.
Caught in the Middle: The Dilemma for AI Vendors
Some may argue that there is no winning for AI vendors. No matter whether they choose to address their models’ biases or not, they will still be criticized. However, the lack of explanation and transparency in the packaging of these models only exacerbates the problem at hand and minimizes the ways in which biases manifest.
Were AI vendors to address their models’ shortcomings head on, in humble and transparent language, it would go a lot further than haphazard attempts at “fixing” what’s essentially unfixable bias. We all have biases, the truth is, and we don’t treat people the same as a result – nor do the models we’re building. And we’d do well to acknowledge that.
Other AI Stories of Note
Here are some other AI stories worth mentioning from the past few days:
AI Models’ Hidden Knowledge
Despite not truly “understanding” what a cat or a dog is, AI models seem to have internalized some “meanings” similar to those of humans. According to Amazon researchers, this is because these models are able to recognize similar “trajectories” between sentences with different grammar but similar concepts. This suggests that these models may have more sophisticated understandings than expected.
Enhancing Prosthetic Vision
Neural encoding has proven useful in improving the limited resolution of artificial retinas and other methods of replacing parts of the human visual system. Using machine learning, Swiss researchers found that they could do perceptual compression, which increases the fidelity of images transmitted to these prosthetics.
“Artificial retinas provide blind people with visual sensations by stimulating the surface of the retina with electrodes. Unfortunately, the resolution is very low and the image relayed to the brain is a mere succession of white dots. However, Dario Floreano and his team from EPFL’s Laboratory of Intelligent Systems have recently been working on an algorithm capable of enhancing visual perception in prosthetic vision. The results are published in Nature Machine Intelligence. ‘We developed a balloon model similar to a saliency map like those used to pay attention when seeing the world,’ he explains. ‘Depending on the subject’s head movements, the balloon inflates and signals an important region to stimulate. In other words, a spectacled subject, who directs his gaze at the world at all times with a rapid succession of small eye movements in all directions, would see a more precise picture than a fixed subject.'”
LLMs Making Strides in Chemistry
Researchers at EPFL have found that LLMs can be fine-tuned and utilized in chemistry with minimal training, proving to be a useful tool in expanding the knowledge of chemists. This is due to its ability to be trained on a large body of work that no individual chemist could possibly know all of. For instance, LLMs can be trained to answer yes or no questions about certain chemical properties, and soon are able to extrapolate from that trained knowledge.
Images vs. Text in Gender Stereotyping
Berkeley researchers have found that images found on Google are much more likely to enforce gender stereotypes for certain jobs than text mentioning the same thing. They also discovered that people are more likely to associate certain roles with one gender when viewing images rather than reading text, even days later.
“This isn’t only about the frequency of gender bias online. Part of the story here is that there’s something very sticky, very potent about images’ representation of people that text just doesn’t have,” explained researcher Douglas Guilbeault.
Despite the advances made in AI, it’s crucial to remember that the source of data for many of these models contains serious biases that have real effects on people.
