Click the image for details↑
Human vision is an extraordinary ability. Although it has evolved over millions of years in specific environments, it can accomplish tasks that early visual systems have never experienced. Reading is a great example, such as recognizing cars, airplanes, road signs, and other man-made objects.
However, the visual system also has a series of significant flaws, namely the optical illusions we experience. In fact, researchers have discovered many ways that can lead people to misjudge colors, sizes, relative positions, and movements.
Illusions themselves are fascinating because they can provide deep insights into the nature of vision and perception. Therefore, exploring these boundaries by discovering new methods of illusions can be greatly beneficial.
Image | Concentric Circles? (Source: MIT Technology Review)
This is where deep learning comes into play. In recent years, computers have learned to recognize objects and faces in images and can create similar images themselves. Therefore, it is easy to imagine that machine vision systems should be able to learn to recognize optical illusions and create illusion images themselves.
Let’s join the research of Robert Williams and Roman Yampolsky from the University of Louisville, Kentucky. These researchers attempted this feat but found that it is not straightforward. Current machine learning systems cannot generate optical illusions on their own—at least not yet. Why is that?
First, let’s look at some background. Recent advances in deep learning are based on the progress of two technologies. The first is the effectiveness of powerful neural networks, which can learn well with just a few programming tricks.
The second is the establishment of large annotated databases that computers can use to learn. For example, to train a computer to recognize faces, it needs thousands of clearly labeled images containing faces. Using that information, a neural network can learn to identify facial features—such as two eyes, a nose, and a mouth. More remarkably, two neural networks—generative adversarial networks—can guide each other to create realistic yet completely synthetic facial images.
Williams and Yampolsky approached the task of training a neural network to identify optical illusions in the same way. Computational power is easily accessible, but the necessary databases are lacking. Therefore, the researchers’ primary task is to create a database of optical illusions for training.
It turns out that this is quite challenging. “The number of static illusion images is only in the thousands, and the unique types of illusion images are certainly very low, perhaps only in the dozens,” they said.
This poses a challenge for current machine learning systems. “Creating a model from such a small and limited database represents a huge leap in model generation and understanding of human vision,” they stated.
Thus, Williams and Yampolsky edited a database containing over 6,000 optical illusion images and subsequently trained a neural network to recognize them. They then created a generative adversarial network to enable the neural network to generate illusions on its own.
The results were disappointing. “After training for 7 hours on an NVIDIA Tesla K80 graphics card, no valuable images were generated,” the researchers said, adding that they made their database available for others to use.
However, this is an interesting outcome. “The only known optical illusions to humans are created by evolution (like the eye patterns on butterfly wings) or by human artists,” they noted.
In both cases, humans play a crucial role by providing important feedback—humans can see the illusions. “Without understanding the principles behind these illusions, generative adversarial networks seem unable to learn how to deceive human vision,” Williams and Yampolsky said.
This is not easy, as there are significant differences between machine vision systems and human vision systems. Many researchers are developing neural networks that are closer to the human visual system. Perhaps whether they can see illusions will be an interesting test.
Meanwhile, Williams and Yampolsky are not optimistic. “It seems that an illusion image dataset may not be sufficient to produce new illusions,” they said. Therefore, optical illusions represent a fortress of human experience that machines have yet to conquer.
-End-
