Scientists propose a solution to address AI bias.

2019-12-22

In recent years, due to some breakthroughs in artificial intelligence technology, computers have learned to observe the world more clearly. However, industry insiders are well aware of the flaws in these artificial intelligence algorithms. A recent experiment has shown that the best artificial intelligence visual system may spit out racial slurs, gender stereotypes, or vocabulary that defames someone's character when seeing a photo. Now, scientists in the field of machine learning vision have eliminated some of the human biases hidden in the data they use for training. They believe that these changes can help artificial intelligence view things more fairly. However, these efforts also indicate that eliminating biases in artificial intelligence systems is still challenging, partly because they still rely on humans to train them. The elimination of artificial intelligence biases mainly aims to eliminate hidden biases and prejudices in automated systems. This is a crucial issue because the deployment speed of artificial intelligence is fast and can have a significant impact. Biases have been confirmed in algorithms behind facial recognition systems, recruitment programs, and web search. Visual systems are being applied in critical areas such as law enforcement, where biases may make surveillance systems more likely to mistake ethnic minorities for criminals. In 2012, a project called ImageNet played a crucial role in unleashing the potential of artificial intelligence. This project provided developers with a massive library for training computers to recognize visual concepts, including everything from flowers to snowboards. Scientists from Stanford, Princeton, and the University of North Carolina paid a small amount of money to the team to label over 14 million images, gradually accumulating a large amount of data. When this dataset is fed into a large neural network, it creates an image recognition system that can identify things with astonishing accuracy. The algorithm learns from many examples to recognize patterns that reveal advanced concepts, such as the texture and shape of a puppy. A competition using algorithms developed with ImageNet showed that the best deep learning algorithms could correctly classify images and people. The success of systems built on ImageNet has sparked a wave of artificial intelligence hype and investment, as well as advancements in other fields, bringing about new technologies such as advanced smartphone cameras and automated vehicles. However, in the following years, other researchers discovered hidden problems in the ImageNet data. For example, if the image library labeled as "programmers" is biased, algorithms trained using this data may assume that programmers are white. Recently, a viral web project called "Mining Artificial Intelligence" also highlighted biases in the labels added to ImageNet. Through this project's website (which is now offline), people could submit a photo and see the terms hidden in the artificial intelligence models trained using the dataset. These exist because the people adding the labels may have included derogatory terms and vocabulary beyond labels like "teacher" or "woman." The ImageNet team analyzed their dataset to uncover these biases from this and other sources, and then took steps to address these issues. The team assessed the population and geographic diversity in the photos and developed a tool to present more diverse images. For example, typically, the word "programmer" might generate a lot of photos of white people in front of computers. But with the new tool that the organization plans to release in the coming months, a more diverse set of images can be generated, showing greater diversity in terms of gender, race, and age, and used to train artificial intelligence algorithms.

RECENT AI NEWS

RECENT AI TOOLS