Blind Spots in Visualizing Artificial Intelligence Models

2019-11-13

With the continuous advancement of technology, GAN (Generative Adversarial Network) has become increasingly skilled at drawing faces. It can predict what a person will look like in the future and can also transform an ordinary person into a celebrity through simulation. However, it is not suitable for drawing large scenes. A new demonstration from MIT reveals that trained models can make decisions on their own when drawing scenes at church and monument sites, deciding what to draw and what to omit. At a recent international computer vision conference, they presented a larger study on the blind spots of GAN. In simple terms, the researchers wanted to know what GAN cannot draw. "Researchers typically focus on describing and improving what machine learning systems can do and what they prioritize. However, we hope that researchers can pay more attention to the data features that these systems ignore," said David Bau, a graduate student at MIT. In GAN, a pair of neural networks work together to create highly realistic images based on given examples. MIT professor Antonio Torralba and IBM researcher Hendrik Strobelt developed a tool early on that made it possible for artificial neurons to organize images into real-world categories such as doors, trees, and clouds, laying a solid foundation for subsequent developments. This led to the creation of image editing software GANPaint, which has become one of the favorite tools for amateur artists. One day, while helping an artist use GANPaint, David Bau noticed a problem. As usual, he was tracking numbers and trying to optimize the photo. However, halfway through, he remembered his mentor's advice: don't just look at the numbers, carefully examine the actual image. When he followed this advice, a strange phenomenon occurred - in large scenes, people were selectively excluded and ignored. This piqued David Bau's interest. Later, David Bau and his colleagues trained different types of GANs in indoor and outdoor scenes. However, regardless of the background of the photo, GAN consistently ignored important details. For example, in a scene of a newlywed couple kissing on the steps of a church, after being reconstructed by GAN, the people disappeared, but the wedding dress remained on the church door. Some researchers suspect that the main reason is machine inertia. Despite being well-trained, GAN tends to ignore certain statistically significant details and skips over the reconstruction of complex images. "Artificial intelligence that removes objects from their representation only achieves its digital goals, while ignoring important details for humans," added David Bau. In short, it can be said that the culprit is not just machine inertia, but also algorithmic issues - the lack of understanding of human aesthetics by machines.

RECENT AI NEWS

RECENT AI TOOLS