A crowd of people in England. Photo: Pictures Ltd./Corbis via Getty Images.

IBM plans to release more than 1 million facial images to help better train the artificial intelligence behind facial recognition systems.

Why it matters: The risk of bias being built into AI systems is a major hurdle for all companies developing facial analysis algorithms to, for example, recognize different skin colors and other attributes in a non-discriminatory way. Since AI is only as good as the data that trains it, IBM thinks making a diverse dataset available will help root out bias.

How it works: IBM says its new project will be the largest facial image dataset available that is specifically curated for the training of AI, and will be open to academics, public interest groups and competitors.

  • IBM selected a sampled subset of one million images from Yahoo's Flickr dataset (Flickr allows the use of its images for research).
  • Its researchers have been annotating those images with facial attributes (such as hair color and facial hair) for more advanced matching capabilities, and they used geo-tags to get an appropriate mix of data from multiple countries.
  • IBM is also releasing a dataset of 36,000 facial images that are equally distributed across all ethnicities, genders and ages to provide a more diverse dataset for researches to use to identify and correct bias in their facial analysis systems.
"No single company can tackle the challenge of AI bias in a vacuum, and we believe it’s essential that tools like this be available for everyone in the field so we all can play a role in advancing the technology responsibly."
— Ruchir Puri, chief architect, IBM Watson

The challenge: AI technologies are developing quickly, but consumers are wary of the consequences. At the same time, consumers are becoming more aware of the power of their data — including photos and videos — being collected by tech giants to build AI systems.

  • For months, IBM has been pushing responsible technology development, making the case that losing consumer trust in the early days of AI will undermine the broader benefits the tech industry thinks AI can offer. Of course, loss of consumer trust is also bad for business.

Go deeper: Axios' Ina Fried has more on the risk of bias in AI:

Go deeper

FDA chief vows agency will not accept political pressure on coronavirus vaccine

Food and Drug Administration Commissioner Stephen Hahn promised that "science will guide our decision" for a coronavirus vaccine at a Senate hearing on Wednesday.

Why it matters: More Americans are expressing doubt about a first-generation vaccine, despite President Trump's efforts to push an unrealistic timeline that conflicts with medical experts in his administration.

CEO confidence rises for the first time in over 2 years

Data: Business Roundtable; Chart: Naema Ahmed/Axios

A closely-watched CEO economic confidence index rose for the first time after declining for nine straight quarters, according to a survey of 150 chief executives of the biggest U.S. companies by trade group Business Roundtable.

Why it matters: The index, which still remains at a decade low, reflects corporate America's expectations for sales, hiring and spending — which plummeted amid uncertainty when the pandemic hit.

Official says White House political appointees "commandeered" Bolton book review

John Bolton's book "The Room Where it Happened." Photo: Chris Delmas/AFP via Getty Images

A former career official at the National Security Council claims her pre-publication review of former national security adviser John Bolton's explosive book on President Trump was "commandeered by political appointees for a seemingly political purpose," according to a letter from her lawyers filed in court on Tuesday.

Why it matters: The White House fought against the publication of Bolton's book for most of the year on the grounds that it contained harmful and "significant amounts of classified information."

Get Axios AM in your inbox

Catch up on coronavirus stories and special reports, curated by Mike Allen everyday

Please enter a valid email.

Subscription failed
Thank you for subscribing!