'Research to increase the resolution of mosaic image by 64 times' developed into a discussion of racial discrimination, researchers who accused of discontinuation of account



Facebook's Chief Researcher in the Artificial Intelligence division, Jan Lucan , who won the 2018 Turing Prize for a person who has achieved achievements in the field of computer science, discussed racism in artificial intelligence and machine learning. Following a lot of criticism, we announced the suspension of our Twitter account.

Yann LeCun Quits Twitter Amid Acrimonious Exchanges on AI Bias | Synced
https://syncedreview.com/2020/06/30/yann-lecun-quits-twitter-amid-acrimonious-exchanges-on-ai-bias/

The trigger began on June 20, 2020, when Duke University published the results of research on high-quality image generation using artificial intelligence algorithms on Twitter.



Duke University announced a technology that converts a 16x16 pixel mosaic-like image into a fine 1024x1024 pixel image in a few seconds. You can see examples of image conversion and the technology used in the article below.

A technology will be developed to increase the resolution of mosaic images by 64 times and produce infinitely high quality images-GIGAZINE



Duke University's research is published on GitHub, and you can actually try image conversion at the following website.

Face Depixelizer Eng-Colaboratory
https://colab.research.google.com/github/tg-bomze/Face-Depixelizer/blob/master/Face_Depixelizer_Eng.ipynb

However, the research conducted by Duke University is not a technique to 'reconstruct' a detailed image from a mosaic image, but a technique in which artificial intelligence infers the original image from the image and 'creates a new image.' In fact, the person who converted the mosaic image of former US President Barack Obama questioned that Obama's face was not restored correctly.



In response to the tweet above, Associate Professor Brad Weibull of Pennsylvania State University said, 'This image shows the danger of prejudice in artificial intelligence.'



And Lecan responded to Weibull's tweet, 'The learning data of the machine learning system is biased. This face upsampling system has been pre-trained on FlickrFaceHQ , which mainly contains white photos. So the conversion results look like everyone is Caucasian, and if you do the exact same training on the Senegal dataset, everyone will look like Africans.'



``I'm tired of this framing,'' said Tim Knit Gebul , founder of the Black in AI group that protects blacks in the field of artificial intelligence, and technical leader of the ethical artificial intelligence team at Google. Many people have tried to explain, and many scholars have tried to explain. Listen to us. Don't just blame the machine learning on the dataset.' did.



Known for his research on race and gender bias in facial recognition systems and artificial intelligence algorithms, Mr. Gebul has advocated fairness and ethics in artificial intelligence for many years and argued that 'Let's diversify machine learning datasets' I've been. A study he did with MIT Media Lab revealed that 'commercial face recognition software is more likely to be misclassified by women with darker skin compared to men with lighter skin'. I am.

Lucan argued that his comments were on a specific dataset used by Duke University. It also states that it is the engineers who use machine learning that need to pay attention to data selection, not researchers who study machine learning.



But Gebul said to Lecan's tweet that he was 'believable' and said, 'We say that people like him should learn. We are trying to educate people in our own community.' It's a tweet.



The approximately one-week discussion between Lecan and Gebul drew thousands of comments and retweets, and many prominent artificial intelligence researchers complained about Lecan's explanation. 'I don't respect Lecan's opinion in a respectful way, as long as machine learning is benchmarked with race-biased data,' said Google Research scientist David Har. It's reflected as a bias. It's useless to ask engineers to retrain with unbiased data in biased machine learning.'

Many people participated in discussions on machine learning and racial issues, including Mr. Lekan and Mr. Gebul, as well as artificial intelligence researchers and activists on racism. In addition to the discussion, many tweets gathered that simply attacked either Lecan or Gebul.

'I highly respect Gebul's work on AI ethics and impartiality,' said Lucan, 25 June 2020. About working to make sure that bias is not amplified by artificial intelligence. I am deeply concerned. I am sorry that the method of communication here has been a topic of conversation.' Tweeted and apologized to Mr. Gebru.



And on 29 June 2020, Lecan said on Twitter, 'Stop posting each other on Twitter and other means, following last week's post. Especially critical of Mr. Gebru and my post. Stop attacking everyone, whether it's words or not, clashes are harmful and counterproductive. I oppose all forms of discrimination. We posted to Facebook about our beliefs and this is our last important post on Twitter,' tweeted and announced that she will stop posting to her account.



Following discussion on Twitter, the research team at Duke University also updated the paper on June 24, 2020. In the paper, 'Overall, sampling from StyleGAN seems to be much more frequent on white faces than on colored faces. The generated photos show racial bias. , 72.6% of the white photos are white, 13.8% are Asian, 10.1% are black, and Indians represent a small percentage of the photos, 3.4%.' ..

[2003.03808] PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models
https://arxiv.org/abs/2003.03808

Related Posts:

in Mobile,   Software,   Web Service,   Science, Posted by darkhorse_log