How OpenAI’s DALL·E 2 illustrated the challenges of bias in AI

An artificial intelligence program that has wowed the internet with its ability to generate original images from user prompts has also sparked concern and criticism for a now-familiar problem with AI: racial and gender bias.

And while OpenAI, the company behind the program dubbed DALL E 2, has attempted to address the issues, the effort has also come under scrutiny because some technologists have claimed it’s a shallow way to solve systemic underlying problems with AI fix systems.

“It’s not just a technical problem. This is a problem that affects the social sciences,” said Kai-Wei Chang, an associate professor at UCLA’s Samueli School of Engineering who studies artificial intelligence. There will be a future where systems are better at protecting against certain biases, but as long as society has biases, AI will reflect that, Chang said.

OpenAI released the second version of its DALL·E image generator in April to rave reviews. The program prompts users to enter a series of words that relate to each other – for example: “An astronaut plays basketball with cats in space in a minimalist style.” And with space and object awareness, DALL·E creates four original images that according to the website the words should reflect.

As with many AI programs, it didn’t take long some user to report what they saw as signs of prejudice. OpenAI used the sample caption “a builder” which produced images showing only males, while the caption only produced “a flight attendant”. images of women. In anticipation of these prejudices, OpenAI released a “risks and limitations‘ document with the program’s limited release before allegations of bias were raised, noting that “DALL E 2 additionally inherits various biases from its training data and its results sometimes reinforce societal stereotypes.”

DALL·E 2 relies on another piece of AI technology called OpenAI GPT-3a natural language processing program that draws on hundreds of billions of language examples from books, Wikipedia, and the open web to create a system that can approximate human writing.

Last week, OpenAI announced that it was implementing new mitigation techniques that have helped DALL E create more diverse and reflective images of the world’s population – and that’s what it claimed Internal users were 12x more likely to say images contained people from diverse backgrounds.

That same day, Max Woolf, a data scientist at BuzzFeed, who was one of a few thousand people granted access to test the updated DALL·E model, began a twitter thread The reference to the updated technology is said to be less accurate than previously when creating images based on his written request.

Other Twitter users who tested DALL·E 2 responded to Woolf’s thread and shared the same issue – specifically regarding ethnic and gender bias. They hypothesized that OpenAI’s diversity solution would be as simple as having the AI ​​append gender or race-identifying words to user-written prompts without their knowledge to generate inorganically distinct sets of images.

“The way this rumored implementation works is that it randomly adds either male or female or black, Asian or Caucasian to the prompt,” Woolf said in a phone interview.

OpenAI published a blog post last month about trying to fix bias by rebalancing certain data; Nothing was mentioned about adding gender or race designations to prompts.

“We believe it’s important to address bias and security at all levels of the system, which is why we’re taking a number of approaches,” an OpenAI spokesperson said in an email. “We are exploring other ways to correct for bias, including best practices for fitting training data.”

Concerns about bias in AI systems have increased in recent years as examples in automated hiring, healthcare, and algorithmic moderation have been found to discriminate against different groups. The topic has sparked debate government regulation. New York passed a corresponding law in December Banned the use of AI to screen job applicants unless the AI ​​passed a “bias audit”.

A big part of the AI ​​bias problem stems from the data that trains AI models to make the right decisions and achieve the desired outcomes. The extracted data often has built-in biases and stereotypes due to societal prejudice or human error, such as B. Photo datasets depicting men as executives and women as assistants.

AI companies, including OpenAI, then use data filters to prevent graphical, explicit or otherwise undesirable results, and in this case images, from appearing. When the training data is passed through the data filter, what OpenAI calls “bias amplification” produces results that are more skewed or biased than the original training data.

This makes it particularly difficult to fix AI bias after a model has been created.

“The only way to really fix this is to retrain the entire model on the biased data, and that wouldn’t be short-term,” Woolf said.

Chirag Shah, an associate professor at the Information School at the University of Washington, said that AI bias is a common problem and that the fix OpenAI appears to have developed has not solved the underlying problems of its program.

“The common denominator is that all of these systems are trying to learn from existing data,” Shah said. “They are shallow and superficial and fix the problem without fixing the underlying problem.”

Jacob Metcalf, a researcher at Data & Society, a nonprofit research institute, said one step forward would be for companies to be open about how they build and train their AI systems.

“For me, the problem is transparency,” he said. “I think it’s great that DALL·E exists, but the only way these systems will be secure and fair is with maximum transparency into how they are being managed.”

https://www.nbcnews.com/tech/tech-news/no-quick-fix-openais-dalle-2-illustrated-challenges-bias-ai-rcna39918 How OpenAI’s DALL·E 2 illustrated the challenges of bias in AI

Fry Electronics Team

Fry Electronics.com is an automatic aggregator of the all world’s media. In each content, the hyperlink to the primary source is specified. All trademarks belong to their rightful owners, all materials to their authors. If you are the owner of the content and do not want us to publish your materials, please contact us by email – admin@fry-electronics.com. The content will be deleted within 24 hours.

Related Articles

Back to top button