A brand new evaluation will (hopefully) make you assume twice about utilizing latent diffusion fashions with out questioning the outputs
We’re seeing main advances in synthetic intelligence arrive thick and quick whereas new units are being launched on a month-to-month foundation. Some AI insiders have expressed concern that it’d trigger human extinction, whereas others say that we shouldn’t be distracted by potential long-term results as a result of the imminent danger is way more vital. On the opposite hand, might synthetic intelligence be the important thing to unlocking Africa’s potential? Or has Joe Biden simply killed the excitement? The AI tides change each week…
Most of the AI information we see considerations Large Language Models (LLMs) equivalent to ChatGPT and Bard. But latent diffusion fashions that create photographs, equivalent to Midjourney and Dall-E, have improved considerably in technical high quality over the previous few months. Can the identical be mentioned for his or her capability to supply a consultant vary on the subject of human faces and locations?
If there’s one space the place machine studying has failed miserably, it’s lowering ethnic stereotyping. I already wrote about my very own rough-and-ready check three months in the past and now there’s a main evaluation by Rest of World that proves simply how reductionist AI picture turbines are. The authors carried out experiments utilizing easy prompts on Midjourney equivalent to “A Nigerian particular person” or “A home in India”. The outcomes are astonishing of their lack of selection. They are additionally uncomfortably and perhaps even dangerously reductionist.
Ethno-cultural stereotyping isn’t the one bias revealed by the experiment. Bizarre ageism and sexism stereotypes are rampant. Apparently, an Indian particular person is in all circumstances an older Hindu man, whereas an American particular person is overwhelmingly white, younger, and handsome.
It seems {that a} narrowing of variety and flattening of options are inherent in machine studying methods whose purpose is to supply a median…
