Our brains can be incredibly deceptive, and it turns out, so can artificial intelligence! This revelation is not only fascinating but also sheds light on the inner workings of our minds.
When we gaze up at the moon, it appears larger near the horizon compared to when it's higher in the sky, even though its size and distance from Earth remain constant. This optical illusion is just one example of how our perception can be misleading.
These illusions are like shortcuts our brains take to extract the most crucial details from our surroundings. But what happens when we present these illusions to AI systems?
AI systems, particularly those powered by deep neural networks, are designed to excel at detecting patterns and details that often elude us. They've proven effective in medical scans, for instance, identifying early signs of diseases.
However, some deep neural networks are susceptible to the same visual tricks as humans. This susceptibility provides a unique opportunity to gain insights into how our brains function.
"Using DNNs in illusion research allows us to simulate and analyze how the brain processes information and generates illusions," explains Eiji Watanabe, an associate professor of neurophysiology.
So, what does this mean for our understanding of optical illusions? Well, it's a bit of a mystery. While there are theories, most illusions still lack a definitive explanation.
Studying individuals who don't experience optical illusions has provided some clues. For example, a person who regained sight after being blind since childhood was not fooled by shape-related illusions but could perceive motion illusions.
This suggests that our ability to perceive motion might be more resilient to sensory deprivation than our ability to make sense of shapes.
Brain imaging studies using fMRI have also revealed which parts of the brain are active when we see different illusions and how they interact. However, our perception of optical illusions is highly subjective, making it challenging to study objectively.
This is where AI steps in as a game-changer. Many AI algorithms, including chatbots like ChatGPT, are powered by deep neural networks that mimic how our brains process information.
In a recent study, Watanabe and his colleagues used a deep neural network called PredNet, which is based on the predictive coding theory of how our brains handle visual information.
They trained the AI using videos of natural landscapes, similar to what humans might see when looking around. Then, they presented the AI with variations of the rotating snakes illusion, a static image that appears to be rotating.
Watanabe found that the AI was tricked by the same images as humans, supporting the predictive coding theory. However, there were also differences in how the AI and humans perceived the illusion.
"This is likely because PredNet lacks an attention mechanism," Watanabe explains. In other words, the AI processes the entire image without focusing on specific spots, unlike humans.
While AI systems can mimic certain aspects of our visual system, they are still far from seeing the world as we do.
Inspired by this work, Ivan Maksymov developed a model that combined quantum physics and AI to simulate how we perceive the Necker cube and Rubin vase illusions.
Maksymov found that the AI regularly switched between interpretations of the illusions, similar to how humans perceive them. This suggests that certain aspects of human thought, like decision-making, can be better modeled using quantum theory.
Such a system could also be used to simulate how our visual perception might change in space under different gravitational conditions.
"While it's a narrow field of research, it's quite important because humans want to go to space," Maksymov says.
As we explore the wonders of the universe, it's crucial to understand how our perception might be affected, ensuring that space travelers can trust their eyes.
