Motion Illusions Generated Using Predictive Neural Networks Also Fool Humans
A new AI model generates novel visual illusions that trick human perception, supporting a key theory of how our brains work.
Researchers Lana Sinapayen and Eiji Watanabe have published a significant paper demonstrating that artificial intelligence can generate novel visual motion illusions that successfully fool human perception. Their system, called the Evolutionary Illusion GENerator (EIGen), is built on a video-predictive neural network and creates static images that appear to move. The work provides compelling evidence for a long-standing hypothesis in neuroscience: that motion illusions are a side effect of the brain's predictive processing capabilities. Essentially, we may sometimes 'see' motion because we are perceiving the brain's own anticipatory signals rather than direct sensory input from the eyes.
The technical achievement lies in EIGen's ability to generate new, validated illusions, moving beyond merely replicating known optical tricks. The researchers confirmed the illusions' effectiveness through a psychometric survey with human participants. Beyond neuroscience, the paper's philosophical contribution is a call to action for AI research to explore 'motivated failures'—designing artificial systems that fail in the same ways biological systems do. This approach could lead to more brain-like AI and offer new tools for understanding human cognition, bridging the fields of artificial intelligence, computer vision, and artificial life.
- The EIGen model uses a video-predictive neural network to generate novel static images that appear to move.
- A psychometric survey confirmed the AI-generated illusions are effective at tricking human visual perception.
- The findings support the predictive processing theory, suggesting illusions occur when we perceive the brain's predictions over raw input.
Why It Matters
This research provides a new tool for neuroscience and points toward building AI that fails in human-like, informative ways.