DALL · E 2 OpenAI creates AI images that are biased or NSFW • Registry

In brief Experts warn that the OpenAI DALL · E 2 image rendering model may be fun to play with, but exhibits common stereotypes, prejudices, and can create NSFW images, making it risky to implement in the real world.

“Without sufficient barriers, models like DALL · E 2 could be used to generate a wide variety of deceptive and otherwise harmful content and could influence how people perceive the authenticity of content more generally,” admitted the newcomer. “DALL · E 2 additionally inherits various biases from the training data and its results sometimes reinforce social stereotypes.”

The Soup Bowl, which is a portal to another dimension of digital art by DALL · E 2

Given a prompt that includes the words “flight attendant”, for example, the model skews the model, creating images of women, while the “builder” shows mostly men. But that’s not the only concern, according to The Guardian.

Trained on images scraped over the Internet, DALL · E 2 can sometimes create works that include logos or trademarks, making some people wonder if this will cause future copyright issues.

DALL · E 2 is currently only available to select developers and artists for testing purposes. OpenAI wants to see how the tool behaves and assess its level of risk before it is released as a commercial API like its GPT-3 text generation system.

Can AI express its understanding of the physical world through text?

MIT researchers trained a neural network to automatically annotate video and audio data in text.

Alexander Liu, a graduate of the Computer Science and Artificial Intelligence Laboratory at an American university, said the project is teaching machines to better understand many data formats.

“The main challenge is how can a machine adapt these different modalities?” Liu said in a statement. “As humans, it’s easy for us. We see a car and then we hear the sound of a car going by and we know it’s the same. But with machine learning, it’s not that simple. ”

Liu and his colleagues trained the system to take raw data and encode sample characteristics into an array of numbers. These data points represent different words that describe a specific action. Trained with a limited number of videos, the model can choose from 1,000 words to describe him. For example, it should correctly mark a movie with the verb “cry” if it detects the sight and sounds of a sobbing child.

AI has no knowledge of the real physical world as long as it is limited to processing one type of data, such as text. Scientists believe that by giving machines the ability to recognize objects and express concepts in language, it will one day help robots better understand their surroundings. You can read the article describing the project here.

Soundcloud launches a musical AI to predict the best hits

Independent music platform Soundcloud has acquired Musiio, makers of machine learning software that promises to automatically listen to songs and predict whether they will become popular or not, announced this week.

Musicians can upload their songs to SoundCloud, and some have managed to get attention and win record deals. The site is less selective than other popular platforms such as Spotify or Apple Music; listeners need to research more to find the music they like. SoundCloud hopes Musiio’s tools will facilitate this process.

“SoundCloud supports more music from more creators than any other platform in the world,” its CEO said in a statement. “The acquisition of Musiio is accelerating our strategy to better understand how this music moves in a proprietary manner, which is critical to our success.”

Musiio built a neural network capable of processing audio tracks to extract the characteristics of a specific track, analyzing things like its tempo or energy to automatically label its genre and sound. These, in turn, can help you automatically select a playlist based on your songs Soundcloud listeners already like helping them discover new artists they may have never met before. ®

Leave a Reply